tumor versus matched-normal sequencing analysis … · tumor versus matched-normal sequencing...
TRANSCRIPT
UNIVERSITY OF CALIFORNIA
SANTA CRUZ
TUMOR VERSUS MATCHED-NORMAL SEQUENCINGANALYSIS AND DATA INTEGRATION
A dissertation submitted in partial satisfaction of therequirements for the degree of
DOCTOR OF PHILOSOPHY
in
BIOMOLECULAR ENGINEERING AND BIOINFORMATICS
by
J. Zachary Sanborn
December 2012
The Dissertation of J. Zachary Sanbornis approved:
Professor David Haussler, Chair
Professor Joshua Stuart
Professor Nader Pourmand
Dean Tyrus MillerVice Provost and Dean of Graduate Studies
Copyright c© by
J. Zachary Sanborn
2012
Table of Contents
List of Figures v
List of Tables xiii
Abstract xiv
Dedication xv
Acknowledgments xvi
1 Introduction 1
2 Previous Work 72.1 UCSC Cancer Genomics Browser . . . . . . . . . . . . . . . . . . . . . . 72.2 BamBam v1.0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3 Methods to analyze sequencing data of tumors and their matched-normals 153.1 Somatic & germline variant calling . . . . . . . . . . . . . . . . . . . . . 15
3.1.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193.2 Small insertions and deletions . . . . . . . . . . . . . . . . . . . . . . . . 203.3 Relative copy number, allele-specific copy number, and allele fraction . . 223.4 Inferring regions of structural variation using paired-end clustering . . . 243.5 Refinement of structural variation using Bridget . . . . . . . . . . . . . 263.6 Bridget de novo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.6.1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303.7 Tumor Genome Browser . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4 Integrating sequencing results to locally reconstruct the tumor genome 384.1 Breakpoint traversal to infer local tumor genome assemblies . . . . . . . 394.2 Breakpoints associated with highly amplified regions . . . . . . . . . . . 42
iii
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424.3.1 Homozygous deletion of CDKN2A in glioblastoma multiforme . . 424.3.2 Reconstructing highly amplified regions using whole genome se-
quencing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464.3.3 FISH confirmation of double minutes . . . . . . . . . . . . . . . . 564.3.4 Evidence of tumor-specific reads in blood . . . . . . . . . . . . . 60
4.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5 Inferring clonality and tumor evolution 665.1 Copy number alterations, allele fraction, and the allelic state diagram . 705.2 Fitting sequencing data to the ASD . . . . . . . . . . . . . . . . . . . . 795.3 Modeling the clonal mixture of a tumor sample . . . . . . . . . . . . . . 825.4 Linking mutations to clone-specific allelic states . . . . . . . . . . . . . . 86
5.4.1 Via direct phasing . . . . . . . . . . . . . . . . . . . . . . . . . . 865.4.2 Via amplified allele fraction . . . . . . . . . . . . . . . . . . . . . 875.4.3 Comparing allele fractions to infer mutation timing . . . . . . . . 88
5.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 905.5.1 Clonality of glioblastoma multiforme tumors . . . . . . . . . . . 915.5.2 Clonality of squamous cell carcinomas of the lung . . . . . . . . . 1005.5.3 Phased mutations to allelic states . . . . . . . . . . . . . . . . . . 102
5.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6 Conclusion 1126.1 Clone-directed mutation discovery . . . . . . . . . . . . . . . . . . . . . 1136.2 Clinical applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1146.3 Final thoughts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
iv
List of Figures
2.1 UCSC Cancer Genomics Browser. (a) Whole-genome view of 119 breasttumor samples measured on a MIP comparative genomic hybridization(CGH) array, sorted by Her2 status (black: Her2-, yellow: Her2+). Am-plifications and deletions are shown in red and blue, respectively, whereasnormal copy number is shown in white. A Student’s t-test was per-formed comparing Her2+ vs. Her2-, red indicates genomic regions thatare highly amplified in the Her2+ group. (b) Zoomed into chromosome 17for both the CGH data shown in (a) as well as gene expression measuredon Affymetrix U133A microarrays for a superset of the breast tumor sam-ples from (a), sorted similarly and with the same t-test applied. Up- anddown-regulated genes are shown in red and green, respectively, whereasno change in expression is shown in black. Note the ERBB2 locus is bothamplified and highly expressed in the Her2+ samples, as expected. Aregion co-amplified with the ERBB2 locus is found upstream on chromo-some 17. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.1 Normalized mutated triplets for (a) 17 GBM tumors and (b) 11 LUSCtumors, sequenced by TCGA. The scale is mutations per million sites,which corrects for the uneven distribution of base triplets in the humangenome. The samples listed on the right provide the order of samplesdisplayed on the bar graphs, from top-to-bottom. . . . . . . . . . . . . . 21
v
3.2 (a) Overview of the Bridget method to refine structural variants detectedby BamBam using split reads. Two sets of 16bp tiles (red and bluelines) derived from the putative split read are shown to align to differentsegments of the reference genome. The two reference segments A & Bare each defined by a “spanning set” of contiguous tiles in the referenceand unaligned read, and the pair of them define the “dual-spanning set”(DSS) that describe the breakpoint junction. Note that the overlap of thetwo regions in the DSS defines the junction’s homology. (b) An exampleBridget result for a tandem duplication identified in a serous ovariancarcinoma sample sequenced by TCGA, denoted OV-0751. A total of 6unaligned reads contributed to this result (sandwiched by the respectiveregions in the reference genome), discovering a 3bp microhomology of“AGC” at the junction point. . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3 Bridget de novo results for 13 whole genome GBMs and their matched-normals. (a) Histogram of indel size showing distinct peaks for insertionsand deletions corresponding to SINE and LINE families of repetitive el-ements. (b) Junction homology for germline (blue) and somatic (red)breakpoints showing peaks that indicate at least two primary mech-anisms for breakpoint induction and repair: microhomology-mediatedend-joining (MMEJ), non-homologous end joining (NHEJ) and repeat-mediated, non-allelic homologous recombination (NAHR). Note the en-richment of MMEJ & NHEJ breakpoints among the somatic breakpoints. 33
3.4 Distance between Bridget-refined breakpoints vs. breakpoints availablein the Database of Genomic Variants (DGV). Each dot represents a singlebreakpoint found by Bridget, where its location on the x-axis is defined byD(bbridget, bdgv) + 1.0 between the Bridget solution and the closest break-point in DGV. A distance of 1.0 is a perfect score to place it on a log scale.The cluster at 2bp is due to a difference in how DGV stored breakpointsbelonging to “Sequence mapping” category, utilizing a one-based genomecoordinate system instead of the zero-based coordinates that Bridget re-ports. Also, not that some studies such as those categorized as “Illuminasequencing” have higher differences, with a peak around 12bp. Thoseparticular studies did not incorporate split read methods to refine theirbreakpoints. Similarly, breakpoints derived from chip-based assays havemuch higher differences from the closest Bridget solution, as expected. . 35
vi
3.5 UCSC Tumor Genome Browser. The browser shows all of the high-levelsomatic difference discovered by BamBam in a single image, enablingthe synthesis of multiple distinct datasets to give the user an overallpicture of the tumor’s genome. The browser is able to zoom into and outof genomic regions rapidly, going from the full genome view, as shownabove, to a single base resolution in just a few clicks. (a) Linear view ofa highly-rearranged and amplified region on chromosome 12 in GBM-06-0648. (b) A circular genome view of the same region showing how theamplified peaks on chromosome 12 are connected to an amplified regionof chromosome 22. It also shows a set of clustered rearrangements onchromosome 9 that caused the homozygous deletion of CDKN2A/B. . . 36
4.1 Toy examples of breakpoint graphs and their solutions. (a) One possibletwo path solution that accounts for all observed breakpoints and tumorcopy number. Note that the deletions in both paths could be swappedto create another equally valid solution to the breakpoint graph. (b) Acircular solution to the breakpoint graph for a highly-amplified region.The circular nature of the solution may indicate the presence of a doubleminute chromosome and/or a homogeneously-staining region (HSR). . . 41
4.2 Browser shots demonstrating how CDKN2A was homozygously deletedin: (a) GBM-06-0188, (b) GBM-06-0145, and (c) GBM-06-0648. Inter-and intra-chromosomal rearrangements caused both copies of CDKN2Ato be deleted in (a) and (b), while a clustered set of intra-chromosomalrearrangements deleted both copies in (c). . . . . . . . . . . . . . . . . 45
4.3 Browser shot of the inter- and intra-chromosomal, highly-supported break-points and copy number observed for TCGA sample GBM-06-0648. . . . 49
4.4 (a) Diagram of the circular solution to breakpoint graph for GBM-06-0648 that accounts for all observed, highly-supported breakpoints, as wellas the solution for the remaining, un-amplified segments of chromosome12. (b) Circular assembly of the circular solution, possibly indicating thepresence of a double minute chromosome. Note the positions of MDM2and the CPM-Novel Exon gene. (c) The linear solution of the remainingcontiguous segments on chromosome 12. (d) RNA-Seq data for GBM-06-0648 that indicates a novel exon replaced the final exon of CPM in thecircular assembly. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.5 Browser shot of the inter- and intra-chromosomal, highly-supported break-points and copy number observed for TCGA sample GBM-06-0152 for (a)chromosome 12 and (b) chromosome 7. . . . . . . . . . . . . . . . . . . . 52
vii
4.6 (a) Diagram of the two circular solutions to breakpoint graph for GBM-06-0152 that accounts for all observed, highly-supported breakpoints. (b)Circular assemblies of the two circular solutions, possibly indicating thepresence of two independent double minute chromosomes. The largercircular solution has two copies of MDM2 and one copy of CDK4, whilethe smaller solution has EGFR. (c) A zoomed-in browser shot of theamplified segment labeled h, showing the distinct changes in copy numberas described in (a). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.7 Browser shot of the intra-chromosomal, highly-supported breakpointsand copy number observed for TCGA sample GBM-06-0145 for chro-mosome 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.8 (a) Diagram of the three paths that together form a potential solutionof the breakpoint graph for sample GBM-06-0145 that accounts for allobserved, highly-supported breakpoints. The location of the gene EGFRis noted. Note that all paths, A, B and C, can create circular solutionsby using the encompassing tandem duplication shown in red to connectthe end of each path to the beginning. The tandem duplication may alsoconnect path A to path B, path C to path B, etc. (b) The effects of eachpath on the EGFR gene, showing that path B creates an oncogenic formof EGFR, EGFRvIII, through specific deletion of exons 2-7. . . . . . . 57
4.9 Fluorescent in situ hybridization (FISH) images for samples used to de-termine the physical state of the amplified regions. (a) FISH image forGBM-06-0648 using probes for MDM2 (red) and the centromere of chro-mosome 12 (green). Note that the improved focus of the image on theright displays the punctuate pattern in red, characteristic of a DM, con-firming that MDM2 is amplified by DM, as predicted. (b) FISH imagefor GBM-06-0145 using probes for EGFR (red) and centromere of chro-mosome 7 (green). Note that the red probes in most cells are clusteredtogether, as opposed to the pattern seen in (a), indicative of an HSR.However, there are cells in (b), such as the bottom-left cell in the rightimage of (b), that display a DM-like pattern similar to that observedin (a). This suggests that, while the dominant pattern indicates EGFRamplification via HSR, the EGFR amplicon may also take the form ofdouble minute chromosomes. . . . . . . . . . . . . . . . . . . . . . . . . 59
4.10 Bar graphs showing breakpoint read support for tumor (black bars) andblood (red bars) for sample: (a) GBM-06-0152, (b) GBM-06-0648, and(c) GBM-06-0145. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
viii
4.11 Browser shot of raw sequencing reads for GBM-06-0648 surrounding theleftmost position of a highly-supported deletion-type breakpoint that ispart of this sample’s putative DM. (a) The position of the deletion-typebreakpoint and an excerpt of the tumor sequencing reads that supportthe breakpoint and the associated amplicon. (b) The sequencing readsfor the matched-normal (blood) sample for GBM-06-0648 showing a setof 9 sequencing reads that support the breakpoint shown in (a). . . . . . 63
5.1 A toy example of the evolution of a tumor starting from the germline,to an initial tumor cell, to a population of major and minor clones thatare sampled by the tumor biopsy. To avoid complicating this example,mutation and copy number passenger events that occurred prior to theemergence of the initial tumor cell were not included. . . . . . . . . . . . 69
5.2 An example of an allelic state diagram (ASD) of simulated data for amonoclonal tumor sample with zero normal contamination, α = 0. Chro-mosomal regions exhibiting different copy number alterations are plottedin different colors. This simulated tumor genome exhibits 6 allelic states:normal (black), single-copy amplification (yellow), hemizygous deletion(red), homozygous deletion (dark green), copy-neutral LOH (CN-LOH,blue), and multi-copy biallelic amplification (cyan). . . . . . . . . . . . . 72
5.3 A set of allelic state diagrams for the same simulated monoclonal tumorgenome shown in Figure 5.2 with different levels of normal contaminant:(a) 0%, (b) 10%, (c) 50%, and (d) 90%. Notice that, as the level ofnormal contamination increases, the distance between allelic states isreduced, negatively affecting our ability to distinguish between differentstates. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.4 Allelic state diagram of the case presented in Figure 5.2 showing somepossible bidirectional and unidirectional transitions between its allelicstates. Note that unidirectional transitions are those that involve irre-versible loss of a parental chromosome. . . . . . . . . . . . . . . . . . . . 75
5.5 Allelic state diagram of the case presented in Figure 5.2 showing howtransitional allelic states are created when the tumor consists of a mixtureof two different clones. In this case, there are two subclones. Clone Ais defined by the original allelic states: yellow (2,1), cyan (5,2), and red(1,0). Clone B alters these states through amplifications and a deletionto produce the allelic states: yellow (2,2), cyan (4,2), and red (2,0). Theyshare the same allelic states of blue (2,0), black (1,1), and green (0,0).The percentages denote the percentage of clone B present in the tumorpopulation, where 0% describes a monoclonal population of clone A, and100% is a monoclonal population of clone B. . . . . . . . . . . . . . . . . 76
ix
5.6 Allelic state diagram of the case presented in Figure 5.2 showing the tran-sitional allelic states produced when allelic states are “skipped,” whichcan occur when the tumor consists of two or more unrelated, or distantly-related clones. Notice that the transitional states are not found on theedges connecting allelic states if the allelic states of the two major clonesdiffer in both majority and minority alleles. . . . . . . . . . . . . . . . . 78
5.7 Allelic state diagrams for two GBM tumors: (a) GBM-06-0145 and (b)GBM-06-0185. The fitted parameters found normal contamination at21.5% and 14.6%, respectively. Notice that in (a) there are only clonalallelic states and no evidence of transitional allelic states, indicating thatGBM-06-0145 is a monoclonal tumor. However, in (b), there are bothclonal states and multiple transitional allelic states. Since the blue andpurple transitional allelic states feature three different mixture percent-ages, this polyclonal tumor must consist of at least three subclones. Theblack X’s plotted in (b) represent “landmark” allelic states that will beused to determine the clonal mixture of GBM-06-0185. . . . . . . . . . 93
5.8 Monoclonal allelic profile for GBM-06-0145. The “Relative Coverage”and “Allele Fraction” displayed at the top of the plot shows both theobserved output by BamBam and the imputed coverage and allele frac-tion generated by rendering the data as the mixture of the single cloneand normal contamination. The comparison of real versus modeled datashows very strong agreement. The clone’s allelic profile below shows themajority and minority allelic states for the tumor genome, showing am-plification of one copy of entire chr7 and chr19, complete loss of one copyof chr10, and arm-level loss of chr9p. . . . . . . . . . . . . . . . . . . . . 95
5.9 Polyclonal allelic profiles for GBM-06-0185. A total of 4 distinct cloneswere identified in this tumor, with clone D determined to be the dominantclone of the population, comprising 42.7% of the tumor sample. All cloneshave single-copy amplifications of chr7, chr19 & chr20, single copy lossof chr10 and chr22, and loss of chr9p in common. Clones B, C, & D allhave deletions in chr6, but clone B’s deletions are focal while clones C& D display arm-level loss of chr6q. Clone D is further distinguished byamplification of the intact copy of chr9. . . . . . . . . . . . . . . . . . . 96
5.10 Polyclonal allelic profiles for GBM-06-0152. Three clones were identifiedin this tumor sample that has an estimated 24.1% normal contamination.All clones share amplifications of chr1, chr19 & chr20, deletions of chr10& chr22, and focal losses of chr12 related to the chromothripsis-like eventthat created two DMs described in the previous chapter. Clones B & Cexhibit amplification of chr7 and deletions of the non-amplified copy ofchr1 as well as chr2, chr3, chr4, chr8, chr13, and chr17. Clone C furtheramplifies the remaining copy of chr8. . . . . . . . . . . . . . . . . . . . . 98
x
5.11 Polyclonal allelic profiles for GBM-06-1086. Four clones were identifiedin this tumor sample that has an estimated 7.5% normal contamination.All clones share amplification of chr21 and deletions of chr9 & chr11p.Clones C & D exhibit significant chromosomal loss, deleting chr1, chr3,chr4, chr5, chr6, chr8, chr10, chr13, chr14, chr15, chr17, chr18, and chr20.The dominant clone D, making up 41.6% of the tumor sample, furtherdeletes the sole remaining copy of chr18 and amplifies chr19. The blackarrows indicate the position of CDK2NA in clones A & B, highlightingthe arrival of the focal deletion of CDKN2A in the latter clone. . . . . . 99
5.12 Allelic state diagrams for two LUSC tumors: (a) LUSC-34-2596 and (b)LUSC-66-2756. The fitted parameters found normal contamination at36.4% and 19.6%, respectively. These tumors are highly polyclonal, withLUSC-66-2756 exhibiting a large number of highly amplified allelic statesand zero single copy loss allelic states, which could have resulted from agenome doubling event. . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.13 Example of seven phased mutations on the dual ASD. A representativeregion on the tumor genome is shown with blue and red horizontal linesrepresenting chromosomes in the tumor mapping to this region, consist-ing of a subregion in the single copy gain allelic state, a subregion inthe “normal” allelic state, a subregion exhibiting CN-LOH, and a subre-gion exhibiting LOH. Mutations 1, 2, and 3 are found in the amplifiedsubregion, where mutations 2 & 3 are majority-phased (red stars) andmutation 1 is minority-phased (blue star). Mutations 4 & 5 are foundin the “normal” allelic state, where mutation 4 is majority-phased andmutation 5 is minority-phased. Both regions exhibiting LOH have onemutation each phased to the sole remaining allele, and thus mutations6 & 7 are majority-phased. The dual ASD below shows where each ofthese mutations would be found, using each mutation’s corrected allelefraction, MAFc, to determine its location along the x-axis. Note the dif-ferent placement of mutations 2 & 3 in the single-copy gain allelic state,where mutation 3, the mutant allele that exists on both chromosomeswith the majority allele, is found near the single copy gain majority al-lelic state. Mutation 2 is found near the single copy gain minority allelicstate, correctly identifying that the mutation exists on only one copy ofthe majority allele and hence is likely to occur after the duplication event.Finally, note that the minority-phased mutations in blue are all found inthe left-half of the dual ASD. . . . . . . . . . . . . . . . . . . . . . . . . 104
xi
5.14 Phased mutations on the dual ASD for tumor GBM-06-0145. 7 regionsare encircled on these plots: (a) majority-phased mutations in an am-plified allelic state, but presenting with MAFc suggesting the mutationis only on one of the two copies, (b) majority-phased mutations in anamplified allelic state with MAFc suggesting mutations are present onboth amplified copies, (c) majority-phased mutations with an allele frac-tion consistent with the LOH allelic state, (d) minority-phased mutationsin the majority-amplified allelic state with MAFc consistent with singlecopy, (e) unphased mutations with MAFc consistent with amplified al-lelic state, and (f) unphased mutations with MAFc consistent with LOHallelic state. LOH has left this cell with only a mutated copy of the geneDOCK8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.15 Phased mutations on the dual ASD for tumor LUSC-34-2596. The twoencircled regions, (a) and (b), show a number of mutations phased tothe majority and minority alleles, respectively, in the balanced amplifiedallelic state (2,2). One majority-phased mutation in NDRG1 is found ina transitional allelic state with matching MAFc. The locations of twomissense mutations, in BRAF & DNMT3A, and one nonsense mutationin TP53 are shown in the unphased plot, placing BRAF in a highlyamplified allelic state, DNMT3A in the “normal” allelic state, and TP53in the CN-LOH state. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
xii
List of Tables
4.1 Types of CDKN2A Loss in 17 TCGA GBMs. . . . . . . . . . . . . . . . 46
5.1 Clonality of 12 GBMs. . . . . . . . . . . . . . . . . . . . . . . . . . . . 915.2 Phased mutations in GBM and LUSC tumors. . . . . . . . . . . . . . . 110
xiii
Abstract
Tumor versus Matched-Normal Sequencing Analysis and Data Integration
by
J. Zachary Sanborn
Cancer research is a study of tragic complexity. The more we discover about this
disease, the more we realize how much we have to learn. It relentlessly unravels new
complications, justifying the difficulty we have had in ridding cancer from the human
experience. Nevertheless, we strive to conquer it. To accomplish this noble goal, we
require tools to understand a patient’s specific tumor in as comprehensive a manner
as possible. The methods presented here aim to provide a complete assessment of a
tumor’s genome, discovering the alterations made to the patient’s germline genome
that may have given rise to cancer. Mutations both large and small are identified
and integrated using various techniques to discover the events and elements specific
to the tumor genome. The methods will be used on whole genome sequencing data of
several tumors, revealing the structural events that occurred in multiple tumors. Double
minute chromosomes in GBM tumors will be discovered computationally and confirmed
experimentally. Finally, techniques that expose tumor clonality will be described and
used to infer whole genome, allele-specific copy number profiles of all major clones
comprising a tumor population from a single tumor biopsy.
xiv
To my beautiful wife,
Shirley Truong,
for her friendship, support, and, most of all, patience through these many years
of graduate school.
xv
Acknowledgments
I would like to thank my advisor David Haussler for all of his advice and encouragement
throughout my long graduate school journey. I thank my committee members Josh
Stuart and Nader Pourmand for all of their advice. Of course, I thank the many members
of the UCSC cancer group, especially Jing Zhu, Sofie Salama, Chris Szeto, Chris Wilks,
Amie Radenbaugh, Mia Griffor, Ting Wang, and the late Fan Hsu. Lastly, I would like
to thank my friends and Five3 Genomics cofounders Steve Benz and Charlie Vaske who
will hopefully help to make the research presented in this thesis actually improve people’s
lives one day. The work has been partially supported by a NIH Training Grant, the
I-SPY INSTINCT project, The Cancer Genome Atlas (TCGA) Pilot and U24 funding
sources, the ARCS Foundation, and Stand Up To Cancer (SU2C) project.
xvi
Chapter 1
Introduction
According to the American Cancer Society, cancer accounts for nearly a quar-
ter of all deaths in America, nearly 600,000 people in 2012, with only heart disease
killing more people per year [1]. Approximately 1 in 2 males and 1 in 3 females will
develop cancer at some point in their life, with the leading incidence sites being lung,
colon, and prostate (or breast) for males (or females). Fortunately, cancer death rates
have been in modest decline since 1990 for both men and women, with the decrease at-
tributed to behavioral changes, such as reduced rates of smoking, earlier detection with
improved screening methods, such as the prostate-specific antigen (PSA) blood-based
test for prostate cancer, and advances in cancer treatment, but there is great room for
improvement. With the advent of precision medicine, significant advances in the treat-
ment of cancer will be possible through improved understanding of both common and
unique features of each patient’s tumor. For example, the particular treatment giving
1
to a women diagnosed with breast cancer is influenced by measurements of the biolog-
ical markers estrogen, progesterone, and HER2 receptors [94, 77]. Also, patients with
chronic myelogenous leukemia (CML) are likely to respond well to treatment with the
tyrosine kinase inhibitor imatinib (Gleevec) if the gene fusion of BCR and the tyrosine
kinase ABL is observed [23, 66].
In order to formulate effective precision cancer treatments, it is vital to un-
derstand how each cancer fulfills the general “requirements” for tumorigenesis. It is
generally accepted that the formation of cancer requires the acquisition of five essential
characteristics, the so-called “hallmarks of cancer”: self-sufficiency in growth signals,
insensitivity to antigrowth signals, avoiding cell death (apoptosis), unlimited replica-
tion, and sustained angiogenesis [35]. To develop metastatic cancer, the tumor must
further acquire an ability to invade other tissues. Two emerging hallmarks, deregula-
tion of cellular energetics and avoidance of immune destruction, as well as two enabling
characteristics, genomic instability and inflammation, were recently added [36]. What
makes cancer such a complex disease to both study and treat is that there are a number
of different ways that tumors acquire these characteristics. Constitutional activation of
BCL2, mutations to TP53, loss of APAF1 function, and alterations to the PI3K-AKT
survival signaling pathway are but a few ways that cancer cells inhibit apoptosis [61].
Occasionally, single alterations provide the tumor with more than one of these essential
characteristics. For example, in non-small cell lung cancer, angiogenesis is regulated
via VEGF (vascular endothelial growth factor) or amplification of EGFR, which also
promotes tumor cell growth [95]. Cancer research is focused on discovering and under-
2
standing the modifications that initiate tumorigenesis and finding treatments that can
target cancer cells more effectively, with reduced impact on healthy cells. The goal as
always: better treatments with less side-effects.
Identifying perturbations in genes or sets of genes that are shared across many
tumors is a powerful method to discover events that may be important to tumorigenesis.
A small region of chromosome 17 was discovered to be amplified in a subset of breast can-
cers, later found to be harboring the growth factor receptor ERBB2 [11]. This subtype
later become one of five clinically-relevant breast cancer subtypes (basal-like, luminal A
& B, ERBB2-overexpressing, and normal-like) defined by a set of differentially-expressed
genes identified in several independent studies using different microarray technologies
[62, 80, 81]. Interestingly, these gene expression signatures can occasionally help identify
specific genomic events, such as the gene expression signature associated with a recur-
rent genomic translocation in acute lymphoblastic leukemia (ALL) [26]. Algorithms
such GISTIC have been applied to tumor copy number data measured on comparative
genomic hybridization (CGH) microarrays to find recurrent amplifications and deletions
in various types of cancer [74, 10, 73]. Four recurring mutations in the coding sequence
of three genes, IDH1, NRAS, and NPM1, and a potential regulatory region were dis-
covered in a large cohort of acute myeloid leukemia (AML) samples [49]. Probably
most famous of all, mutations in TP53 are nearly ubiquitous in many different types of
cancer, reflecting its general importance to tumorigenesis [33].
Structural rearrangements of the genome are also known to be recurrent events
in cancer, resulting in gene fusions that characterize particular cancers. In many CML
3
cancers, a genomic translocation between chromosomes 9 and 22 (denoted t(9;22)) re-
sults in the fusion of two genes, BCR and ABL, to form the chimeric protein BCR-ABL
[8]. When localized to the nucleus, ABL can induce apoptosis in response to DNA dam-
age, but the chimeric protein BCR-ABL is localized exclusively to the cytoplasm and
activates anti-apoptotic pathways [91]. By performing a meta-analysis of several gene
expression datasets on prostate cancer, researchers discovered chimeric proteins created
by fusing TMPRSS2 to members of the ETS family (ERG and ETV1), one of the first
recurrent gene fusion events found in cancers of epithelial origin [87]. Although recur-
rent structural variations appear to be rare in non-hematological cancers, researchers
argue that this is due to the difficulty in observing these events in these cancers due
to their more complex karyotypes [41]. In fact, it is believed that gene fusions are
equally common in blood, soft, or solid cancers, with the bias due to the relative lack
of recurrent chromosomal aberrations observed in solid cancers [53].
The identification of recurrent events is essentially a search for events that
drive the development and expansion of cancer, so called driver mutations that provide
a competitive advantage in the cellular environment. Recurrent mutations are thought
to be likely drivers because of the low likelihood that similar, or identical, mutations
would occur independently in multiple tumors unless they contributed to tumorigenesis.
This assumes that mutations arise randomly across the genome (i.e. no mutational
hotspots). Cells that acquire these driver mutations are also likely to carry passenger
mutations that provide neither benefit nor detriment to the tumor cell, but which tend
to accumulate in the tumor genome simply because the passenger mutations occurred
4
in cells undergoing tumorigenesis.
Distinguishing which events act as drivers is difficult, especially since recur-
rent events are not always necessarily drivers. Recent studies have attempted to distin-
guish somatic mutations driving the development of cancer from passenger mutations.
Large-scale resequencing of the FLT3 coding region in 222 AML patients identified nine
non-synonymous mutations, of which only four were gain-of-function mutations found to
constitutively activate FLT3’s kinase domain [27]. The remaining neutral mutations ob-
served were likely passengers of an unobserved oncogenic mutation. Similarly, another
study on 188 AML samples found that, out of 30 novel non-synonymous mutations
observed in 22 different kinases, only mutations to FLT3 had any effect on cell trans-
formation [48]. In a study of drivers of liver cancer, researchers used RNAi in a mouse
model to knock out orthologous genes recurrently deleted in human hepatocellular car-
cinomas, identifying 12 tumor suppressors out of the 301 genes tested [96]. Clearly,
the evidence supports the notion that a significant fraction of observed mutations, even
those occurring in the same genes across many tumors, are likely passenger mutations.
To this end, researchers have developed mutational models that better model the non-
uniform distribution of mutations across the genome in order to improve our ability to
discover significantly mutated genes from genes that carry many passenger mutations
[17, 57].
Ultimately, tumorigenesis must be understood from the context of evolution,
where tumorigenesis is the process by which cells adapt in order to survive, and unfor-
tunately prosper, inside a human host that actively seeks to destroy misbehaving cells.
5
Adaptations come about via both loss- and gain-of-function mutations, providing the
emerging cancer cell with the ingredients essential for malignant growth. During the
cell’s evolution towards cancer, an initial tumor cell emerges and rapidly expands to
form the majority of the observed tumor mass. During its growth, additional mutations
may occur that can further drive or impede the cancer’s growth, or are simply benign
passenger mutations. From this evolutionary perspective, driver mutations are those
that provide the tumor cell with a set of adaptations that optimize its ability to grow
within its specific cellular environment. Driver mutations are more likely occur early
in the tumor’s development. This evolutionary process has been studied directly by
investigating the evolution of the Ig heavy chain locus in 22 B-cell chronic lymphocytic
leukemias, enabling researchers to divergent subclones in a phylogenetic tree that rep-
resents the evolution of the germline genome to that of the dominant clone [13]. The
arrival of whole-genome sequences for tumors and matched-normal samples from TCGA
provides an opportunity to not only identify genomic events (mutations, structural re-
arrangements, copy number) at a very high resolution but, as will be shown shortly,
chronologically order genome-wide events to understand a single cancer’s evolution [56].
6
Chapter 2
Previous Work
My disseration is based upon my previous work and the work of other members
of the Haussler cancer group, which begins with work focused on the visualization and
analysis of microarray-based, high-throughput data and ended with my first foray into
whole genome sequencing analysis of tumor & matched-normal samples.
2.1 UCSC Cancer Genomics Browser
Given cancer’s heterogenous nature, it is important to acquire as much in-
formation on each tumor in order to better understand the state of each individual
cancer. To this end, cancer studies like TCGA use microarrays to measure gene expres-
sion, copy number variation, and methylation on each tumor and, when appropriate,
matched-normals derived from blood or solid tissue of the same patient. TCGA has
also identified somatic and germline variants through targeted gene re-sequencing using
7
traditional Sanger sequencing, and has just begun whole-genome sequencing of tumors
using next generation sequencing platforms. By combining this data across hundreds of
different tumors within the same tumor type, researchers hope to find genomic features
that characterize a specific type, or subtype, of tumor while also noting the unique char-
acteristics of each tumor. After collecting this variety of data on thousands of tumors,
it becomes very important to have a tool that aids in the exploration and analysis of
these data.
We developed the UCSC Cancer Genomics Browser (CGB) as a tool specifi-
cally designed for these large cancer studies [98, 67]. It is designed to simultaneously
explore many types of data, both genomic and clinical, measured on multiple cancers in
a single web-based platform accessible to both bioinformaticians and cancer biologists.
The CGB interface features two heatmaps, one displaying genomic data in either genome
coordinates or “genesets” and another displaying any available “clinical” data pertain-
ing to each sample displayed. These two heatmaps are coupled vertically, meaning that
each row of both heatmaps describes data pertaining to the same tumor sample. Within
this vertical coupling lies much of the power of the CGB, because, with a single click,
a user can sort samples based on a clinical feature and immediately visualize any pat-
terns in the genomic heatmap that results. For example, amplification/over-expression
of receptor ERBB2/Her2 characterizes all Her2+ breast cancer patients, a fact that
becomes immediately apparent after sorting the array CGH data of a cohort of breast
cancer tumors by their Her2 status, as shown in Figure 2.1. Interestingly, Figure 2.1
shows the presence of another amplicon just upstream in 17q11 that appears amplified
8
in the same tumors with ERBB2 amplification. These two distinct amplicons have been
observed before in many Her2+ breast tumors, and may result by translocation of one
region to the other followed by amplification of the newly adjacent regions as a single
amplicon [11, 19, 43]. The CGB also features a set of simple statistical tools, such as
Student’s t and Wilcoxon rank-sum tests, to help discover associations between genomic
and clinical data that are not as readily visible.
Another major feature of the UCSC CGB is its ability to group together ge-
nomic data into genesets, instead of being restricted to plotting data in genomic co-
ordinates. Since genes often interact with genes that may be located anywhere in the
genome, it would be very difficult to analyze genomic data in the proper biological
context if constrained to genome coordinates, as the UCSC Genome Browser does.
The CGB’s “genesets view” does away with genomic coordinates by displaying the ge-
nomic data of related genes side-by-side in genesets. Over 7,000 genesets were collected
from various databases, such as MSigDB, Reactome, Kyoto Encyclopedia of Genes
and Genomes (KEGG), BioCarta, Gene Ontology (GO), and National Cancer Institute
Pathway Interaction Database (NCI-PID), and consist of genes associated by cellular
processes, genes discovered to be up- or down-regulated in certain cellular conditions,
functional gene modules derived from computational scans of high-throughput protein-
protein interactions, and other methods [85, 90, 38, 4, 68]. In addition to the built-in
genesets, users can create their own genesets. All genesets are searchable via geneset
name or by the name of any genes they contain.
The UCSC Cancer Genomics Browser is still under active development by
9
Her2ER PR pCR
Her2ER PR pCR
Her2ER PR pCR
I-SPY Copy Number (MIP Array, 119 Breast Cancer Samples)
I-SPY Gene Expression (Affy U133A, 121 Breast Cancer Samples)
a.
b.
ERBB2 locus
Co-amplified region
Figure 2.1: UCSC Cancer Genomics Browser. (a) Whole-genome view of 119 breasttumor samples measured on a MIP comparative genomic hybridization (CGH) array,sorted by Her2 status (black: Her2-, yellow: Her2+). Amplifications and deletions areshown in red and blue, respectively, whereas normal copy number is shown in white.A Student’s t-test was performed comparing Her2+ vs. Her2-, red indicates genomicregions that are highly amplified in the Her2+ group. (b) Zoomed into chromosome 17for both the CGH data shown in (a) as well as gene expression measured on AffymetrixU133A microarrays for a superset of the breast tumor samples from (a), sorted similarlyand with the same t-test applied. Up- and down-regulated genes are shown in red andgreen, respectively, whereas no change in expression is shown in black. Note the ERBB2locus is both amplified and highly expressed in the Her2+ samples, as expected. A regionco-amplified with the ERBB2 locus is found upstream on chromosome 17.
10
the cancer genomics group led by Jing Zhu, with numerous improvements having been
made. In addition to the recent upgrades to the web application’s user interface, the
team have added significant capabilities to the browser, including the ability to dynam-
ically generate sample-specific metrics based on mathematical combinations of gene
expression and/or copy number data, enabling gene-based biomarkers to be developed
and immediately tested on the many thousands of samples available on the CGB.
2.2 BamBam v1.0
Near the end of August 2009, TCGA released whole-genome sequencing data
for GBM and ovarian serous carcinoma tumors and matched-normal data derived from
blood of the same patients. The data is stored in the BAM format, the block-compressed,
binary equivalent of the Sequence Alignment/Map (SAM) format [46], which has be-
come the de facto standard for storing aligned single- and/or paired-end short-reads
due to its ease of use, quick read/write access, and well-designed software libraries. The
original GBM tumor and normal samples were sequenced on Illumina GA sequencer to
produce approximately 4-5 billion 76bp paired-end reads for each sample. The ovarian
tumor and normal samples were sequenced on a later model Illumina GAII sequencer,
providing roughly 3 billion paired-end 100bp reads. All reads were aligned to the hg18
reference using MAQ [47].
To determine if a mutation discovered in a tumor sample is somatic or a
germline variant requires that we analyze both the tumor and matched-normal data.
11
This can be done sequentially, by summarizing data at every genomic position for both
tumor and matched-normal and later combining the results for analysis. Unfortunately,
because whole-genome BAM files are hundreds of gigabytes in their compressed form
(1-2 terabytes uncompressed), the intermediate results required for later analysis will be
extremely large and slow to merge and analyze. To avoid this and improve the efficiency
of analyzing these large datasets, I wrote the first version of my sequencing analysis tool
called BamBam.
What made BamBam unique among other sequencing analysis methods of its
time was that it performed all of its analyses by simultaneously reading both tumor
and matched-normal BAM files, piling up reads that overlap every common genomic
location between the two files. This gave BamBam the ability to perform complex
analyses, such as paired tumor vs. matched-normal mutation calling, inferring copy
number alterations, that required sequencing data from both tumor and normal very
quickly and with negligible memory use. Another important benefit to processing these
files in sync was that the output BamBam produced could be drastically minimized since
only the output could be restricted to just the likely somatic differences, producing what
is essentially a whole-genome diff between the tumor and its germline.
The initial version of BamBam incorporated rudimentary consensus somatic
SNP-calling, germline and somatic inter- and intra-chromosomal rearrangements, es-
timates of relative copy number, and finding regions exhibiting loss-of-heterozygosity
(LOH).
Consensus SNP-calling is performed by a weighted-majority vote of the ac-
12
ceptable reads in the pileup, weighted by their Phred-scaled base quality scores:
MV (x) = argmaxθ
N∑i
δ(θ = bi, qi > qmin)qi, (2.1)
where θ ∈ {A,G,C, T}, δ is equal to 1 if its arguments are true, and bi and qi are the
i-th nucleotide base and quality scores of the N reads piling up at the current genomic
location, x, respectively. If the consensus base call in the tumor is different than in the
matched-normal, then a SNP variant is called for that location. For LOH analysis, if
there are enough acceptable reads to call a second allele at the location in the matched-
normal, but no reads with that second allele present in tumor, then LOH is called for
that location. For example, if the tumor has just 5 reads calling A and germline has
4 reads calling A and 3 reads calling G, then this location may have lost the G allele
in the tumor via deletion. By considering only this single location, it could be that
the second allele mutated from G→A in the tumor or that the G allele was simply not
sampled. However, if we instead consider all of the LOH calls within the same genomic
neighborhood and find the density of locations experiencing LOH is high relative to the
average region, then it becomes more likely that the set of heterozygous alleles was lost
through a deletion event rather than multiple locations all experiencing specific base
substitutions or due to sampling bias.
BamBam v1.0 identified putative intra- and inter-chromosomal structural vari-
ants by clustering “discordant” paired reads where each read in the pair map to dis-
parate regions of the hg18 reference. The discordant pairs that meet user-defined quality
thresholds are aggregated to determine the two genomic regions that may be newly adja-
13
cent in both tumor and matched-normal. Discordant reads are clustered if they overlap
on both sides of the putative rearrangement. To mitigate the effect of low coverage of
the breakpoint, such that discordant reads do not directly overlap each other, a 500bp
buffer is placed around the start and stop positions of each discordant read to extend
its genomic span. When at least two pairs of reads overlap in this way, a structural
variant is called and the breakpoint that defines it is recorded in the BamBam out-
put for post-processing and visualization purposes. This version of the rearrangement
detection algorithm only allowed a single discordant connection to exist at any given
genomic location, so it could miss structural variants that are nearby (within 500bp) or
directly overlap other structural variants.
14
Chapter 3
Methods to analyze sequencing
data of tumors and their
matched-normals
The methods described below are available in a pair of programs I wrote: an
updated version of BamBam & Bridget. Both of these programs are designed to run
quickly and use little memory, making them suitable to run on single-core and multi-core
machines in addition to high-performance clusters.
3.1 Somatic & germline variant calling
Because BamBam keeps the sequence data in the pair of files in sync across the
genome, a complex mutation model that requires sequencing data from both tumor and
15
matched-normal BAM files as well as the human reference can be implemented easily.
This model aims to maximize the joint probability of both the germline genotype (given
the matched-normal reads and the reference nucleotide) and the genotype of the tumor
(given the germline genotype, a simple mutation model, an estimate of the fraction of
contaminating normal tissue in the tumor sample, and the tumor sequence data).
To find the optimal tumor and germline genotype, we aim to maximize the
likelihood defined by
P (Dg, Dt, Gg, Gt|α, r)
= P (Dg|Gg)P (Gg|r)P (Dt|Gg, Gt, α)P (Gt|Gg) (3.1)
where r is the observed reference allele, α is the fraction of normal contamination, and
the tumor and germline genotypes are defined by Gt = (t1, t2) and Gg = (g1, g2) where
t1, t2, g1, g2 ∈ {A, T,C,G}. The tumor and matched-normal sequencing data are defined
as a set of reads Dt = {d1t , d2t , ...} and Dg = {d1g, d2g, ...}, respectively, with the observed
bases dit, dig ∈ {A, T,C,G}. All data used in the model must exceed user-defined base
and mapping quality thresholds.
The probability of the germline alleles given the germline genotype is modeled
as a multinomial over the four nucleotides:
P (Dg|Gg) =n!
nA!nT !nG!nC !
n∏i
P (dig|Gg), (3.2)
where n is the total number of matched-normal reads at this position and nA, nT , nC , nG
are the reads supporting each observed allele. The base probabilities, P (dig|Gg), are
16
assumed to be independent, coming from either of the two parental alleles represented
by the genotype Gg, while also incorporating the approximate base error rate of the
sequencer. The prior on the germline genotype is conditioned on the reference base as
P (Gg|r = a) = {µaa, µab, µbb}, (3.3)
where µaa is the probability that the position is homozygous reference, µab is het-
erozygous reference, and µbb homozygous non-reference. The germline prior does not
incorporate any information on known SNPs, to avoid unnecessarily biasing the mu-
tation caller at polymorphic sites. SNPs identified by BamBam as mutations can be
found and filtered by comparing against recent builds of dbSNP [75].
The probability of the set of observed tumor alleles is again defined as multi-
nomial
P (Dt|Gt, Gg, α) =n!
nA!nT !nG!nC !
n∏i
P (dit|Gt, Gg, α), (3.4)
where the probability of each tumor allele is a mixture of base probabilities derived
from both tumor and germline genotypes that is controlled by the fraction of normal
contamination, α,
P (dit|Gt, Gg, α) = αP (dit|Gt) + (1− α)P (dit|Gg) (3.5)
and the probability of the tumor genotype is defined by a simple mutation model from
on the germline genotype
P (Gt|Gg) = max[P (t1|g1)P (t2|g2), P (t1|g2)P (t2|g1)], (3.6)
17
where the probability of no mutation (e.g. t1 = g1) is maximal and the probability
of transitions (i.e. A → G,T → C) are four times more likely than transversions (i.e.
A→ T, T → G). All model parameters, α, µaa, µab, µbb, and base probabilities P (di|G)
for the multinomial distributions can be defined by the user.
The tumor and germline genotypes, Gmaxg , Gmaxt , selected are those that max-
imize Equation 3.1, and the posterior probability defined by
P (Dg, Dt, Gmaxg , Gmaxt |α, r)∑
i,j P (Dg, Dt, Gg = i, Gt = j|α, r)(3.7)
can be used to score the confidence in the pair of inferred genotypes. If the tumor and
germline genotypes differ, the putative somatic mutation(s) will be reported along with
its respective confidence.
Maximizing the joint likelihood of both tumor and germline genotypes helps
to improve the accuracy of both inferred genotypes, especially in situations where one
or both sequence datasets have low coverage of a particular genomic position. Other
mutation calling algorithms, such as MAQ and SNVMix, that analyze a single sequenc-
ing dataset are more likely to make mistakes when the non-reference or mutant alleles
have low support [47, 29]. Since the development of BamBam, new mutation callers
that incorporate both tumor and matched-normal sequencing data have been developed,
further evidence that this is a powerful way to discover somatic variants from tumor
and matched-normal sequencing data [42, 63].
18
3.1.1 Results
Algorithms used to assess the statistical significance of mutations, such as
MutSig [17, 57] and MuSiC [21], normally require 100s of samples. As this thesis
focuses on small numbers of whole genomes, assessing BamBam’s mutation calls from a
significance standpoint is difficult. However, we can instead perform a simple analysis
on the genomic context of mutations called in the whole genomes of 17 glioblastoma
multiforme (GBM) tumors and 11 squamous cell lung carcinomas (LUSC), both by
sequenced by TCGA, to discover if different mutational processes are acting in these
two tumor types.
The genomic context of a mutation consists of the bases that precede and
follow the mutated base in the reference genome. The mutational context is then fully
defined by a mutated triplet, i.e. ACC→ATC, describing a C→T transition within the
A-C genomic context. To avoid undercounting mutations, the base that is mutated is
allowed to be either C or A. If the base is G or T, then the triplet and mutant base are
reverse-complemented. Thus, a total of 32 triplets represent the reference sequence and
can be mutated one of 3 ways, for a possible 96 mutations across all possible triplets.
The total number of mutations in each triplet category are tallied for a tumor or a set
of tumors. These counts are then corrected by the relative frequency of each triplet in
the human genome, to determine if certain genomic contexts are preferentially mutated.
Figure 3.1 shows the total number of somatic mutations found in each triplet
for 17 GBMs and 11 LUSC tumors, normalized by the triplet count in the human
19
reference, in a layout inspired by Mike Lawrence of the Broad Institute (Unpublished).
GBM tumors almost exclusively exhibit CG→TG mutations, one of the most common
mutation types found in the human genome caused by deamination of 5’methyl-cytosine
at methylated-CpG sites [3].
LUSC tumors, on the other hand, display a radically different mutational sig-
nature, exhibiting a preponderance of mutations in almost all triplet contexts. Notably,
C→A (equivalently, G→T) mutations are particularly enriched in all but one LUSC
tumor. This agrees with the known mutational signature of smoking-associated lung
tumors that are characterized by G→T mutations resulting from the repair of bulky
DNA adducts, where polycyclic aromatic hydrocarbons (PAH), the primary carcinogen
found in cigarette smoke, covalently bond to DNA [22, 32]. One LUSC tumor, LUSC-
21-1078, exhibits the fewest mutations of the 11 LUSC tumors, and very few G→T
mutations, and is also the patient that smoked the least out of this set of tumors (9
pack-years vs. median 66 pack-years) [34].
3.2 Small insertions and deletions
BamBam currently employs a simplistic model for calling insertions and dele-
tions (indels). The collection of reads that overlap a position is scanned for any and
all evidence of both insertions and/or deletions. For each deletion, its size and genomic
location is stored. For each insertion, the position of the insertion and the inserted base
sequence is stored. An indel is then called if the identical indel (in location and size
20
CT A G
C
T
A
G
CT A GCT A G
C
T
A
G
prec
edin
g ba
se
48.7 mutations per million sites
C→AC→GC→TA→CA→TA→G
b.
CT A G
C
T
A
G
CT A GCT A G
C
T
A
G
prec
edin
g ba
sefollowing base
71.8 mutations per million sites
a. mutation
C→AC→GC→TA→CA→TA→G
mutation
GBM-16-1063GBM-14-1459GBM-06-0155GBM-14-1401GBM-14-1454GBM-14-0786GBM-06-0877GBM-06-1086GBM-26-1438GBM-16-1460GBM-06-0145GBM-06-0214GBM-06-0152GBM-06-0188GBM-06-0648GBM-06-0185GBM-06-0881
LUSC-66-2756LUSC-60-2722LUSC-34-2600LUSC-34-2596LUSC-56-1622LUSC-66-2757LUSC-60-2713LUSC-43-3394LUSC-60-2695LUSC-60-2711LUSC-21-1078
Figure 3.1: Normalized mutated triplets for (a) 17 GBM tumors and (b) 11 LUSCtumors, sequenced by TCGA. The scale is mutations per million sites, which correctsfor the uneven distribution of base triplets in the human genome. The samples listed onthe right provide the order of samples displayed on the bar graphs, from top-to-bottom.
21
or sequence) is observed more than a user-defined threshold Nindel, where Nindel ≥ 2.
If the indel is found only in the tumor and there exist no indels in the same region of
the matched-normal, then it’s classified as a somatic indel, otherwise it’s classified as
germline. Base quality of the inserted sequence is not incorporated in this model. The
same mapping quality thresholds used for calling point variants are used to filter out
poorly aligned reads.
Using the above model, BamBam currently calls a great number of indel vari-
ants and, from a brief study of these indel calls, it appears that many are likely false
positives produced by systematic errors in the sequencer or aligner. Future versions of
BamBam will incorporate an error model for insertions and deletions, but until such a
model can be developed and tested, indel variants will only be utilized within a limited
scope.
3.3 Relative copy number, allele-specific copy number, and
allele fraction
Relative copy number of tumor vs. matched-normal is estimated using a dy-
namic windowing approach that expands and contracts the window’s genomic width
according to the read coverage in either the tumor or matched-normal data. The pro-
cess is initialized with a window of zero width. Each unique read from either the tumor
or matched-normal sequencing data will be tallied into tumor counts, Nt, or matched-
normal counts, Nn. The start and stop positions of each read define the window’s
22
region, expanding in width as new reads exceed the boundaries of the current window.
When either the tumor or matched-normal counts exceed a user-defined threshold, the
window’s size and location are recorded, as well as the Nt, Nn, and relative coverage NtNn
.
Tailoring the size of the window according to the local read coverage will create wide
genomic windows in regions of low coverage (e.g. repetitive regions), or small windows
in regions that are amplified in the tumor, thereby increasing the resolution of amplicon
boundaries.
Allele-specific copy number is estimated similarly, except that all positions
deemed homozygous in the germline are ignored. Germline heterozygosity is defined as
a position where the majority allele represents from 50-60% of the total number of unique
reads mapping to that position. Majority and minority copy numbers are calculated
separately in a manner identical to that described above for overall copy number. All
counts ascribed to the majority allele in both tumor and matched-normal data will
go towards calculation of the majority copy number, and similarly for the minority
allele. Allele-specific copy number is used to identify genomic regions exhibiting loss-
of-heterozygosity as well as amplifications or deletions specific to one or both parental
chromosomes. This last point is especially important to help distinguish if potentially
disease-causing alleles, such as an inherited TP53 mutant, are retained or amplified in
the tumor genome. Furthermore, regions that experience hemizygous loss (e.g. one
parental chromosome arm) can be used to directly estimate the amount of normal
contaminant in the sequenced tumor sample, which can be used to improve the modeling
of the germline and tumor genotypes described above. This will be described in further
23
detail below.
In most cases, it is more useful to calculate allele fraction since its calculation
does not depend on sequencing data from the matched-normal. Majority and minority
allele fraction is simply defined as:
AFmaj(tmaj , tmin) =tmaj
tmaj + tmin(3.8)
AFmin(tmaj , tmin) =tmin
tmaj + tmin
= 1−AFmaj (3.9)
where tmaj and tmin are the read depths for the majority and minority alleles in the
tumor sample, as defined above. As before, allele fractions are only calculated for
positions that are heterozygous in the matched-normal.
3.4 Inferring regions of structural variation using paired-
end clustering
Putative intra- and inter-chromosomal rearrangements are discovered using
discordant paired reads stored in a pair of coordinate-sorted BAM files, one from the
tumor sample and the other from the matched-normal sample, where each read in
the discordant pair map to disparate regions of the human reference genome. Intra-
chromosomal discordant pairs are those that have an abnormally large insert size (i.e.
the genomic distance on the reference separating the paired reads exceeds a user-
defined threshold) or those that map in an incorrect orientation (i.e. inversion). Inter-
24
chromosomal discordant pairs are defined by paired reads that map to different chro-
mosomes.
All discordant paired-end reads from both tumor and matched-normal datasets
are clustered according to their genomic locations to define an approximate genomic
region where the breakpoint is believed to be. The aggregation process consists of
grouping together the reads that overlap other reads on both sides of the putative
breakpoint and have the identical orientation, while keeping track of the number of reads
that came from tumor and normal datasets. When the number of overlapping discordant
pairs in a cluster exceeds a user-defined threshold, the breakpoint that describes the
rearrangement is defined and recorded in the output.
Breakpoints are classified as “somatic” if they are significantly supported by
reads from the tumor dataset. A minimal amount of read support for “somatic” break-
points from the matched-normal dataset is allowed to accommodate cases where some
minute level of tumor DNA is present in the matched-normal sample. This may oc-
cur when highly amplified regions present in a solid tumor’s DNA are shed off into
the bloodstream at low, but detectable levels. Alternatively, in the case of hemato-
logical cancers, it is expected that there will be a high level of tumor introgression in
the matched-normal sample (typically skin tissue). The amount of support for somatic
breakpoints allowed in the matched-normal dataset can be adjusted according to the
expected amount of contaminating tumor tissue present in the matched-normal sample.
“Germline” breakpoints are those that have significant support in both tu-
mor and matched-normal datasets or only find support in the matched-normal dataset.
25
These breakpoints will not be considered in the analyses presented here.
3.5 Refinement of structural variation using Bridget
The breakpoints discovered from paired-end clustering are approximate such
that they utilize fully-mapped reads that, by their nature, cannot substantially overlap
the actual junction of the two regions that are believed to rearranged. The base sequence
of their junction represents sequence not present in the reference or, in the case of a
somatic structural variant, in the matched-normal dataset. To refine the location of the
breakpoint, we use a program I developed called Bridget that utilizes the sequence of
unmapped reads anchored near the breakpoint to assemble the junction’s sequence. An
overview of this method and an example result is shown in Figure 3.2.
The regions surrounding each breakpoint are searched for all unaligned reads
anchored near the putative breakpoint by a mapped mate. Each of these unmapped
reads has the potential to be a “split read” that directly overlaps the rearrangement’s
breakpoint junction. Localized genomic sequences surrounding both sides of the break-
point are broken up into a set of unique tiles of length 16bp. A database storing the tile
sequences and their location in the reference genome is created. A similar tile database
is constructed for each unaligned read, by breaking up the read into overlapping 16bp
tiles and noting their position within the read. Comparing the reference tile database
and the unaligned tile database, the location of each unaligned tile in the human refer-
ence genome is determined. We then look for sets of tiles that are contiguous in both
26
the reference and the read, defining what is termed a “spanning set” of tiles that span
an equal amount of sequence in both reference and read space. The “dual-spanning
set” (DSS) is a pair of such spanning sets, with one spanning set for each side of the
breakpoint in the reference.
The genomic location of the DSS in reference coordinates precisely determines
the breakpoint’s location and its orientation (or strandedness) on both sides of the
junction. The strandedness is simply determined by looking at each spanning set and
determining if the reference tiles are moving in the same direction (same strand) or
different directions (opposite strands). With the information describing position and
orientation of both left and a right sides of the breakpoint, the junction sequence is
fully defined through a precise rearrangement of the reference genome. The DSS also
describes the sequence homology present at the breakpoint by noting any overlap of
the two spanning sets within the read. Typically, this is short homologous sequence
(microhomology), such as “CAA,” that is identical in the reference sequences on the
both sides of the breakpoint, but is observed only once in the unmapped read where
the two junction sequences overlap.
For each unmapped read processed for a given breakpoint, the DSS deter-
mines a potential location for the breakpoint. Since every unmapped read may de-
termine slightly different locations for the breakpoint (due to sequence errors near the
breakpoint, repetitive reference, etc.), the breakpoint location for every DSS is used
to generate a putative junction sequence. All unmapped reads are realigned using a
Smith-Waterman alignment algorithm to each putative junction sequences and the over-
27
all improvement in their alignments is measured relative to the best alignment found
in the unaltered reference sequence [79]. The junction sequence that yields the greatest
improvement in Smith-Waterman alignment scores is judged as the best candidate for
the true junction. If this best junction sequence yields negligible improvement in the
alignment scores, then no suitable junction sequence could be found. In this case, the
lack of split read confirmation by Bridget could be evidence that the initial breakpoint
discovered by BamBam’s paired-end clustering method was spurious.
3.6 Bridget de novo
Bridget is normally used to simply refine the breakpoints already discovered by
BamBam. To reduce the number of structural variants produced, BamBam only reports
breakpoints found in the expected orientation if the breakpoint’s span (i.e. the distance
between the two sides of the breakpoint) is greater than 2,000bp. Unfortunately, this
means that deletions smaller than 2,000bp will not be found by BamBam. The need
for such a high threshold is due to the fact that the number of called breakpoints goes
up dramatically as the threshold is lowered, thereby greatly increasing the number of
breakpoints that are sent to Bridget for split read confirmation. This is an inefficient
and computationally expensive method for searching for small deletions.
To mitigate this limitation, a secondary mode for running Bridget was devel-
oped. This method, called Bridget de novo, fuses BamBam’s breakpoint caller together
with the Bridget’s split read aligner. This enables breakpoint searching and split read
28
TGCACCCTATATTGTCTGAGAACAGAGTGGCTACACAGAAAATGGAGGCCCACAAAGGGCCACTTCCCCACCTCCCCTCCCTGCCCACTGGCTCCCTCCT
SNP
Putative Split Read Sequence:
Overlap indicates Homology at Junction
kmers constructed from reference (k = 16)
Inferred Junction Sequence:reference segment A reference segment B
...TCTTCCCTCCTCCGAGTGTATAGTGCCATACTCACTCAGGTACCCAAGCtagagacagcagtcagttgtctttgggcctccctatcacatcctataa...
...TCTTCCCTCCTCCGAGTGTATAGTGCCAAACTCACTCAGGTACCCAAGCCTGGCCAACATGGCCAAATCCCATCTATACT GTCTTCCCTCCTCCGAGTGTATAGTGCCATACTCACTCAGGTACCCAAGCCTGGCCAACATGGCCAAATCCCATCTCTACTAAAAACACAAAAATTAGCC TCGTCCCCCCTCCGAGTGTGTAGTGCCATACTCTCTCAGGTACCCAAGCCTGGCCAACATGGCCAAATCCCATCTCTACTAAAAACACAAAAATTAGCCAG GCCATACTCACTCAGGTACCCCAGCCTGCCCCAGGTGGCTAAAACCCCACTCTAGTGGAAGGGCACGGGGTCGCAACGG... GCCATACTCACTCAGGTACCCAAGCCTGGCCAACATGGCCAAATCCCATCTCTACTAAAAACACAAAAATTAGCCAGGG... AATACTCACTCATGTACCCAAGCCTGGCCAACATGGCCAAATCCCATCTCTACTAAAAACACAAAAATTAGCCAGGG... ...gcgggcggatcacctgaggtcaggaattcgagaccAGCCTGGCCAACATGGCCAAATCCCATCTCTACTAAAAACACAAAAATTAGCCAGGG
chr2 : 74,748,737
chr2 : 74,114,535
OV-0751Split Reads
a.
b.OV-0751
Tandem Duplication:
Figure 3.2: (a) Overview of the Bridget method to refine structural variants detectedby BamBam using split reads. Two sets of 16bp tiles (red and blue lines) derivedfrom the putative split read are shown to align to different segments of the referencegenome. The two reference segments A & B are each defined by a “spanning set”of contiguous tiles in the reference and unaligned read, and the pair of them definethe “dual-spanning set” (DSS) that describe the breakpoint junction. Note that theoverlap of the two regions in the DSS defines the junction’s homology. (b) An exampleBridget result for a tandem duplication identified in a serous ovarian carcinoma samplesequenced by TCGA, denoted OV-0751. A total of 6 unaligned reads contributed tothis result (sandwiched by the respective regions in the reference genome), discoveringa 3bp microhomology of “AGC” at the junction point.
29
confirmation to be run as a cohesive pair, where immediately upon calling a putative
breakpoint, split read evidence for that breakpoint is collected and analyzed. Only
breakpoints that are called and have split read support are reported. This enables the
threshold for breakpoint span to be significantly lowered, as the number of false positive
calls are drastically reduced by the need for all calls to be confirmed by Bridget’s inline
split read aligner.
The first step of Bridget de novo analyzes one million aligned read pairs of the
input BAM file to compute statistics on the distribution of insert sizes (i.e. the distance
that separates two read pairs, defined during the preparation of the sequencing library).
It then sets the threshold for breakpoint span to be one standard deviation away from
the computed mean. Thus, the insert size distribution of the sequencing library, instead
of a rigid threshold, determines which breakpoints will be analyzed. For example, for
typical libraries with an insert size of 500±100bp, the lower limit threshold for reads in
the expected orientation will be approximately 600bp. Every breakpoint that exceeds
this minimum threshold, or found in a non-reference orientation (e.g. inverted), and is
also found to have both split read evidence and exhibits significant improvement in its
Smith-Waterman alignment versus reference will be reported by Bridget de novo.
3.6.1 Results
Bridget de novo was independently run on the whole genome sequencing data
of 13 TCGA GBMs and their matched-normals. The two sets of results were merged to
identify the somatic status of each breakpoint, where breakpoints found in the matched-
30
normal exclusively or breakpoints in both matched-normal and tumor were classified
as “germline” and those found solely in the tumor were classified “somatic.” Two
breakpoints are deemed similar if their distance
D(b1, b2) =√
(b1,start − b2,start)2 + (b1,stop − b2,stop)2 (3.10)
is less than 10bp, where b1 and b2 are the two breakpoints being compared and bi,start
and bi,stop are the start and stop positions, respectively, of the putative breakpoint bi.
Figure 3.3(a) shows the distribution of insertion and deletion sizes identified
in these 26 whole genome samples (13 tumors and 13 matched-normals). Peaks at
indel sizes of 320bp and 6,060bp are observed, which match the approximate size of
the SINE and LINE retrotransposons, respectively [64]. Since these mobile elements
are still active, we expect the location and existence of these retrotransposons to be
different in every person compared to the reference genome. Thus, the observation of
these peaks suggest that Bridget de novo is identifying true structural variants in these
genomes. Note that the peak at 320bp could not be identified by BamBam, and thus
Bridget de novo has improved the detection of small structural variants, as desired.
Further evidence of the performance of Bridget de novo is found in Fig-
ure 3.3(b), which shows the distribution of the sizes of the homologous sequenced shared
by the two sequences joined together. The results are shown for germline and somatic
breakpoints separately. Two peaks, at 0-2bp and 14-16bp, are observed in the set of
germline breakpoints. These two peaks suggest that at least two different mechanisms
are primarily responsible for producing the observed breakpoints in the germline. There
31
are three major mechanisms known to induce or repair breakpoints. Microhomology-
mediated end-joining (MMEJ) and non-homologous end joining (NHEJ) are two path-
ways in eukaryotic cells responsible for repairing double-strand breaks (DSBs) [51, 9, 54].
These mechanisms require little to no homology in the sequences to be joined, so they
are likely candidates for the mechanisms responsible for breakpoints making up the 0-
2bp peak. The third major mechanism, non-allelic homologous recombination (NAHR),
requires extensive homology to join two sequences, and is likely the mechanism respon-
sible for producing the indels comprising the 14-16bp peak in Figure 3.3(b) [18].
The distribution of somatic breakpoints in Figure 3.3(b) shows an enrichment
for breakpoints created by the MMEJ or NHEJ mechanisms. This suggests that DSB
repair is a frequent occurrence in the tumor genome, and supports the notion that DSB
induction and repair is an engine of structural change in tumor genomes, as noted by
multiple studies [39, 25].
For more evidence that Bridget de novo is discovering real structural variants,
the breakpoints it identifies are compared to those in the Database of Genomic Variants
(DGV) [37]. The DGV is a collection of germline structural variants discovered by
multiple groups employing a variety of different wet-lab and computational techniques.
Since DGV only stores germline variation, only germline-classified Bridget breakpoints
are compared to DGV. For each germline breakpoint, the closest breakpoint in DGV
is found using the absolute distance metric given in Equation 3.10. The results are
shown in Figure 3.4 for each of the 13 categories of variants available in DGV. This
figure demonstrates that Bridget’s solutions are very close, and often identical, to the
32
Junction Homology (bp)
Count
0-2bp
14-16bprepeat-mediated NAHR
NHEJ, MMEJ
Indel Size (bp)
Count
SINEs320bp
LINEs 6060bp
bridget de novo (min 150bp)bambam (min 2kb)a.
b.
Figure 3.3: Bridget de novo results for 13 whole genome GBMs and their matched-normals. (a) Histogram of indel size showing distinct peaks for insertions and deletionscorresponding to SINE and LINE families of repetitive elements. (b) Junction homologyfor germline (blue) and somatic (red) breakpoints showing peaks that indicate at leasttwo primary mechanisms for breakpoint induction and repair: microhomology-mediatedend-joining (MMEJ), non-homologous end joining (NHEJ) and repeat-mediated, non-allelic homologous recombination (NAHR). Note the enrichment of MMEJ & NHEJbreakpoints among the somatic breakpoints.
33
structural variants identified through sequencing-based methods (top 4 categories). The
array-based variants tend to be more distant from their nearest Bridget breakpoint (i.e.
D(bbridget, bdgv) >> 10bp), but this should be expected as the accuracy of structural
variants inferred from an array is limited by the density and separation of its probes on
the reference genome.
Similarly, the sequencing-based variants in DGV that are distant (>> 10bp)
from the nearest Bridget solution can be explained by the fact that DGV does not require
structural variants to have base-level precision. Therefore, DGV accepts structural
variants identified by paired-end clustering methods like BamBam’s structural variant
caller, which are shown to have reduced precision compared to methods, such as Bridget,
that incorporate split reads.
In both of the above cases where the two solution are distant, the nearest
breakpoints identified by Bridget de novo to the sequencing- or array-based counterparts
stored in DGV are likely refinements of the variant provided by DGV. In most cases,
the Bridget de novo solutions appear to improve the knowledge of the size and location
of these indels in the human reference to base-level precision.
3.7 Tumor Genome Browser
To visualize all of the results output by BamBam, a tumor genome browser
was developed that simultaneously displays all of the genomic variants found in a single
tumor sample, versus its matched-normal, as shown in Figure 4. The browser was built
34
Difference in Breakpoint Locations, | bridget - DGV | (bp)
DGV Category:
Numbers in black = # breaks with dist <= 1000Numbers in yellow = # breaks with dist > 1000
Figure 3.4: Distance between Bridget-refined breakpoints vs. breakpoints available inthe Database of Genomic Variants (DGV). Each dot represents a single breakpointfound by Bridget, where its location on the x-axis is defined by D(bbridget, bdgv) + 1.0between the Bridget solution and the closest breakpoint in DGV. A distance of 1.0 isa perfect score to place it on a log scale. The cluster at 2bp is due to a difference inhow DGV stored breakpoints belonging to “Sequence mapping” category, utilizing aone-based genome coordinate system instead of the zero-based coordinates that Bridgetreports. Also, not that some studies such as those categorized as “Illumina sequencing”have higher differences, with a peak around 12bp. Those particular studies did not in-corporate split read methods to refine their breakpoints. Similarly, breakpoints derivedfrom chip-based assays have much higher differences from the closest Bridget solution,as expected.
35
using the same codebase that powers the UCSC Genome Browser. I wrote a custom
drawing engine and MySQL database designed for rapid extraction and display of all
BamBam results for a single patient. It is capable of displaying overall & allele specific
copy number, intra- and inter-chromosomal rearrangements, and mutations and small
indels. It displays data in both linear and circular genome plots, the latter being much
better suited for displaying inter-chromosomal rearrangements.
a. b.
Figure 3.5: UCSC Tumor Genome Browser. The browser shows all of the high-levelsomatic difference discovered by BamBam in a single image, enabling the synthesis ofmultiple distinct datasets to give the user an overall picture of the tumor’s genome.The browser is able to zoom into and out of genomic regions rapidly, going from the fullgenome view, as shown above, to a single base resolution in just a few clicks. (a) Linearview of a highly-rearranged and amplified region on chromosome 12 in GBM-06-0648.(b) A circular genome view of the same region showing how the amplified peaks onchromosome 12 are connected to an amplified region of chromosome 22. It also shows aset of clustered rearrangements on chromosome 9 that caused the homozygous deletionof CDKN2A/B.
By displaying the data together in a single image, the user can quickly navigate
36
a single sample’s data and understand the relationship between changes in its copy
number and local structural variation. For example, a large intra-chromosomal, deletion-
type rearrangement should have a concordant drop in copy number in the region between
the breakpoint ends. Also, displaying mutation data with copy number data allows the
user to understand if a somatic mutation was subsequently amplified, or if the wild-type
allele was deleted in the tumor, both vital datapoints that may suggest the importance
of the event in this sample’s tumorigenesis.
The ability to visualize and navigate a patient’s data within the genome
browser has been instrumental to many of the discoveries presented in this thesis. While
the browser is still actively used by few members of the cancer group at UCSC, it is
such a valuable tool that it was one of the first things we recreated (and drastically
improved) at my startup, Five3 Genomics, which is used to produce all browser images
presented below.
37
Chapter 4
Integrating sequencing results to
locally reconstruct the tumor
genome
As the tumor genome browser demonstrates, the true power of sequencing
data comes from when the various data, structural rearrangements, copy number, and
mutations, are thoughtfully integrated to form a more complete understanding of the
tumor genome. Through this understanding, we can begin to piece together the events
that characterize the genesis and subsequent development of individual tumors.
38
4.1 Breakpoint traversal to infer local tumor genome as-
semblies
Similar to recently published methods, we construct a breakpoint graph by
describing a set of edges that represent the amplified segments of the tumor genome
and a set of directional vertices that connect the edges to one another [30, 50]. Here, the
edges are defined as the amplified segments of the tumor genome observed in the relative
copy number, while the vertices are the highly supported breakpoints found in the
manner described above. If an amplified segment is interrupted by a breakpoint, then
that segment will be split into two edges at the location of the interrupting breakpoint.
With the segments laid out according to genomic position, we determine the
arrangement of edges that represent the rearranged tumor sequence by starting at the
leftmost position of the first amplified segment and progress towards the right until a
right-oriented vertex is encountered. The path continues by following the vertex to the
segment it is connected to and moving in the direction (left or right) specified by the
outgoing vertex. A complete solution to the breakpoint graph is made when a path
through all edges and vertices have been traversed at least once, satisfying the observed
copy number.
Given such loose constraints, it is clear that many satisfactory solutions to the
breakpoint graph may be possible. However, the optimal path(s) through the graph will
be the one(s) whose inferred copy number most closely agrees with the observed relative
copy number produced by BamBam. The number of times a solution traverses a given
39
segment produces an inference of that segment’s copy number. The root mean square
deviation (RMSD) of the segment traversal counts to the observed relative copy number
for each solution is calculated, and then the solution(s) with the smallest RMSD value
are labeled as optimal.
Two toy example breakpoint graphs and their respective solutions are shown
in Figure 4.1. Both examples in Figure 4.1 shows a set of breakpoints, the relative copy
number between tumor and matched-normal, and respective solutions to the breakpoint
graphs. The first example is a simple one, with four breakpoints and three discrete
copy number states ranging from homozygous deletion to single-copy gain. The solution
presented for this example incorporates two paths that utilize all breakpoints and explain
the observed copy number. However, this is not the only solution to this breakpoint
graph, as the smaller of the two blue deletions could instead be phased with the inverted
breakpoints with no penalty.
The second example in Figure 4.1 describes a slightly more complicated break-
point graph where a clustered set of five breakpoints lie on the boundaries of regions
that are “extremely” amplified, i.e. relative copy number ≥ 10. Here, the simplest
solution to this breakpoint graph involves a circular path through the five breakpoints
and amplified regions. To account for the extra copies observed in the copy number
data, the circular path must be traversed a total of nine times.
In Section 4.3, we describe solutions to breakpoint graphs from real GBM
tumors with qualities similar to both of the examples presented in Figure 4.1.
40
2
1
0
3 copies
Path #1
Path #2
10
10 copies
Path #1
Path #2 (x9)
a.
b.
Highly Amplified Segments
Homozygous Deletion
Observed Breakpoints
Observed Breakpoints
Tumor CopyNumber
Tumor CopyNumber
Figure 4.1: Toy examples of breakpoint graphs and their solutions. (a) One possibletwo path solution that accounts for all observed breakpoints and tumor copy number.Note that the deletions in both paths could be swapped to create another equally validsolution to the breakpoint graph. (b) A circular solution to the breakpoint graph for ahighly-amplified region. The circular nature of the solution may indicate the presenceof a double minute chromosome and/or a homogeneously-staining region (HSR).
41
4.2 Breakpoints associated with highly amplified regions
The support of a given breakpoint is directly proportional to the copy number
of the regions it connects. Thus, by requiring breakpoints to have a high level of read
support, we can filter out breakpoints that are part of copy-neutral rearrangements
or lead to low-copy amplifications and deletions, and instead focus on the breakpoints
that are likely linked to the highly amplified regions in the tumor. The particular read
support threshold is chosen such that breakpoints that have read support expected of
copy-neutral regions of the tumor genome are removed.
4.3 Results
The true test of the efficacy of any computational method is how it performs
on real data. Here we demonstrate the use of the methods described above on real
tumor whole genome sequencing data, creating precise, localized genome assemblies
that integrate the observed copy number and breakpoints produced by BamBam. From
these assemblies, we can infer the structural changes that occurred and possibly discover
the events instrumental to the tumorigenesis of these cancers.
4.3.1 Homozygous deletion of CDKN2A in glioblastoma multiforme
In Figure 4.1, two types of breakpoint graphs, one linear and another circular
were presented. This section presents examples of the first type of breakpoint graph in a
set of glioblastoma multiforme (GBM) tumors and their matched-normals sequenced by
42
TCGA. Using the breakpoints and copy number identified by BamBam and Bridget, we
focus our study on the homozygous deletions of tumor suppressor CDKN2A, a common
occurrence in cancer and notably GBM tumors [60, 56]. CDKN2A inhibits CDK4,
arresting cell-cycle progression into G1 and G2 phases. Tumor cells that delete both
copies of CDKN2A have vastly increased abilities to proliferate [58, 69].
The chromosomal regions containing CDKN2A are shown for three GBM tu-
mors in Figure 4.2. In each case, the structural events by which the copies of CDKN2A
were deleted are different. In Figure 4.2(a) and (b) both show one copy of CDKN2A
deleted through a focal intra-chromosomal rearrangement, while the other copy was loss
via arm-level deletion that resulted from the translocation of the downstream arm to
chromosomes 11 and 22, respectively. The case presented Figure 4.2(c) is complicated
by a greater number of intra-chromosomal rearrangements, but it appears that one copy
was focally deleted via an intra-chromosomal deletion (upper-left deletion breakpoint
colored blue) similar to that presented in (a) and (b). The other copy was deleted
via by the complicated rearrangement formed by the remaining breakpoints. The clus-
tered pattern of breakpoints and oscillating nature of copy number and allele fraction
is suggestive of chromothripsis, an event where catastrophic genomic disruption creates
highly rearranged regions [82, 40].
Table 4.1 lists the status of both copies CDKN2A in a set of 17 TCGA GBM
tumors that we have whole genome sequencing data. 12 of the 17 GBMs experienced
homozygous loss of CDKN2A, while 5 GBMs retained both copies. The general pattern
of loss, shared by all but one GBM sample, is that one copy of CDKN2A was loss via a
43
small focal deletion and the second copy was deleted via a much more dramatic deletion,
i.e. loss of chromosomal arm chr9p, complete loss of chr9, or a set of chr9p deletions
caused by a chromothripsis-like event. This pattern shared by multiple tumors allows
us to speculate on how these events may have transpired in these tumors.
Similar to how we use recurrent mutations to infer oncogenic importance, the
recurrence of focal and large-scale losses of CDKN2A might suggest an importance in
the relative timing of the two events. With this assumption, there are two possibilities:
(1) focal followed by large-scale loss or (2) large-scale followed by focal loss. In both
cases, the loss of the first copy of CDKN2A results in cells with reduced abilities to
stop proliferation, helping these cells survive to the next generation. If we assume
then that these burgeoning tumor cells are under reduced pressure to maintain their
genomic stability, that either loss will be tolerated at first. However, considering that
cells become inviable if they lose both copies of “housekeeping” genes, if a cell loses
its first copy of CDKN2A via loss of the arm or all of chromosome 9, it cannot lose
the second copy by the same means. Only a subsequent focal loss of CDKN2A will
produce viable cells that survive to form the tumor. The same argument can be made
for the focal losses occurring first, but since we see so few examples of double focal losses
suggests that focal losses are less likely to occur first. Of course, all of the above is pure
speculation, without a method to directly infer the timing of the loss of CDKN2A, we
cannot know which came first. The next chapter of this thesis will describe just such a
method.
44
CDKN2A/B
CDKN2A/B
CDKN2A/B
a.
b.
c.
GBM-0188
GBM-0145
GBM-0648
Figure 4.2: Browser shots demonstrating how CDKN2A was homozygously deletedin: (a) GBM-06-0188, (b) GBM-06-0145, and (c) GBM-06-0648. Inter- and intra-chromosomal rearrangements caused both copies of CDKN2A to be deleted in (a) and(b), while a clustered set of intra-chromosomal rearrangements deleted both copies in(c).
45
Table 4.1: Types of CDKN2A Loss in 17 TCGA GBMs.
Sample CDK2NA Copy #1 CDKN2A Copy #2
GBM-06-0145 focal loss chr9p loss via translocationGBM-06-0188 focal loss chr9p loss via translocationGBM-06-0155 focal loss chr9p loss via translocationGBM-06-0185 focal loss chr9p loss via translocationGBM-06-0214 focal loss chr9p loss via translocationGBM-14-0786 focal loss chr9p lossGBM-14-1401 focal loss chr9p lossGBM-16-1063 focal loss chr9p lossGBM-06-1086 focal loss complete loss of chr9GBM-06-0208 focal loss complete loss of chr9GBM-06-0648 focal loss chromothripsis-related focal lossGBM-06-0877 focal loss focal lossGBM-06-0152 No loss No lossGBM-06-0881 No loss No lossGBM-14-1454 No loss No lossGBM-16-1460 No loss No lossGBM-26-1438 No loss No loss
4.3.2 Reconstructing highly amplified regions using whole genome se-
quencing
Next, we will apply the above methods on three GBM tumors from the TCGA
project, GBM-06-0648, GBM-06-0152, and GBM-06-0145, and attempt to reconstruct
regions in each tumor that are found to be extremely amplified. In each case, we find
circular solutions to their breakpoint graphs that could indicate the presence of double
minute chromosomes (DMs), self-replicating, circular chromosomes that are often found
in high copy number in tumors, conferring drug resistance and harboring oncogenes drive
tumorigenesis and growth [7, 65, 70, 7, 6]. By using Bridget-refined breakpoints, the
circular assemblies presented in this section are accurate to the base-level.
46
Double Minute Assembly in GBM-06-0648
A total of 3,696 breakpoints were identified in GBM-06-0648, of which 132
breakpoints were found to have split reads directly spanning the putative breakpoint.
Using a minimum read support threshold of 100 filtered all but 16 of the somatic break-
points with split read support were removed. As shown in Figure 4.3, each of the
remaining 16 breakpoints is proximate to the boundaries of highly amplified segments
in a clustered region of chromosome 12 and a small 250bp region on chromosome 9. A
total of 16 amplified segments are found, total size approximately 891kb, where one seg-
ment contains the known oncogene MDM2 implicated in GBM tumorigenesis. In fact,
the edges of every amplified segment can be uniquely associated with a single breakpoint
oriented such that it enters into or exits from the amplified region. This suggests that the
highly supported breakpoints and amplifications represent the rearranged configuration
of this region in the tumor’s genome.
Figure 4.4(a) represents the single optimal path through the amplified segments
discovered by walking the breakpoint graph. The interesting aspect of this solution is
that it is circular, Figure 4.4(b), where the final traversal of last segment returns back
to the initial position, i.e. the first vertex is also the final vertex. This circular solution
passes through every amplified segment exactly once, yet the copy number suggests that
the average tumor cell contains multiple copies of each segment. This high-multiplicity
circular solution suggests two possibilities: (1) multiple copies of these 16 segments are
inserted into the tumor genome in tandem such that each copy maintains the precise
47
order and orientation described by the circular solution (e.g. homogeneously-staining
region, or HSR), or (2) a single 891kb double minute chromosome containing one copy
of each of the 16 segments may have been formed of which the average tumor cell
has accumulated multiple copies. Note that these two possibilities are not mutually
exclusive, as highly amplified regions are known to exist in a state of flux alternating
between HSRs and DMs [84].
The assembly of the rearrangements specific to the contiguous, non-DM-associated
segments of chromosome 12 is shown in Figure 4.4(c), meaning that the entire struc-
ture of chromosome 12 can be explained by the breakpoints and copy number results
produced by BamBam and Bridget. The observation that deletions surround some of
the amplified peaks suggest that these regions were excised from the genome by some
process (e.g. chromothripsis) and circularized to form the double minute, in what is
termed as the “episome model” of double minute creation that is confirmed in both
hematological and solid tumors [14, 83, 84]. The putative DM contains the oncogene
MDM2, which represses the tumor suppressing activity of TP53 [16].
In the circular assembly of GBM-06-0648, we observe that one of the DM-
associated breakpoints is located within the final intron of the gene CPM. We may
expect that a rearrangement interrupting a gene would severely impact the expression
of its transcript. Fortunately, the Broad Institute had sequenced the RNA of this
same tumor, providing us the ability to measure the expression of CPM. Here we find
something unexpected, not only do we find that CPM is expressed, we find that reads
mapped to the penultimate exon are connected to a region that is directly downstream
48
of CPM only in the circular assembly of GBM-06-0648’s double minute. This provides
yet further support of the veracity of the circular assembly, as this region is 1,470kb
distant from the CPM exon in the human reference, but only 13.5kb away in the circular
assembly. The truly interesting aspect to this discovery is that the region connected to
the penultimate exon of CPM is a completely novel exon, complete with a canonical
splice site and terminating stop codon, birthed out of normally unexpressed intergenic
DNA. A BLASTp search of this novel exon’s 25 amino acid protein sequence did not
find any significant hits, so the functional impact of this novel exon to the activity of
CPM is currently unknown [2].
MDM2
Figure 4.3: Browser shot of the inter- and intra-chromosomal, highly-supported break-points and copy number observed for TCGA sample GBM-06-0648.
Assembly of two independent double minutes in GBM-06-0152
A total of 1,669 somatic breakpoints were identified in GBM-06-0152, of which
120 breakpoints were found to have split reads directly spanning the putative break-
point. 22 breakpoints surpassed the minimum read support threshold of 100 to identify
49
chr12
chr9
(~225 bp)w
a b dc e f h i k l m n o p q r s t u vjg
b.
CPM-Novel Exon
MDM2
CAND1RAPB1(Partial)
chr9
*
ACFLAFCSRKELGIDWKQFGGIDIWVLVSS
CPM, Exon 8
10bp
d. c r
Novel Exon
*
•AG
*chr12a
chr12e -u j -g v
c.
891kb~50 copies
a.
Figure 4.4: (a) Diagram of the circular solution to breakpoint graph for GBM-06-0648that accounts for all observed, highly-supported breakpoints, as well as the solution forthe remaining, un-amplified segments of chromosome 12. (b) Circular assembly of thecircular solution, possibly indicating the presence of a double minute chromosome. Notethe positions of MDM2 and the CPM-Novel Exon gene. (c) The linear solution of theremaining contiguous segments on chromosome 12. (d) RNA-Seq data for GBM-06-0648that indicates a novel exon replaced the final exon of CPM in the circular assembly.
50
breakpoints likely associated with highly-amplified regions, including two breakpoints
that did not have split read evidence. All 22 of these highly supported breakpoints are
proximate to the boundaries of 29 highly amplified segments in clustered regions on
chromosome 7 and chromosome 12, as shown in Figure 4.5, suggesting that the highly
supported breakpoints and the amplifications are directly associated and may represent
the rearranged configuration of one (or multiple) amplicons in the tumor’s genome.
Figure 4.6 shows the solution to the breakpoint graph of GBM-06-0152 that
predicts two independent paths that completely account for all observed breakpoints
and amplified segments. Path A has a total length of 1,269kb, uses 18 of the 20 intra-
chromosomal breakpoints on chromosome 12, and contains one copy of CDK4 and two
copies of the oncogene MDM2. CDK4 interacts with CDKN2A to control cell-cycle
progression in the G1-S phase, and is known to enhance tumor progression when over-
expressed [72, 52]. Path B has a total length of 929kb, incorporating the remaining
two intra-chromosomal breakpoints on chromosome 12 and the two inter-chromosomal
breakpoints to connect the amplified segments on chromosome 12 to the amplified seg-
ment containing EGFR on chromosome 7. Both paths are circular in nature and suggest
that two independent oncogenic DMs were created and amplified in this tumor. The
observation that the EGFR-DM contains two amplified segments within a cluster of
rearrangements specific to the MDM2-CDK4-DM suggests that the creation of these
two DMs were coincident.
To incorporate all highly supported breakpoints in the MDM2-CDK4-DM,
multiple traversals of some segments were required. 11 of the 29 segments were traversed
51
a.
b.
Amplified germline breakpoint
MDM2CDK4
Figure 4.5: Browser shot of the inter- and intra-chromosomal, highly-supported break-points and copy number observed for TCGA sample GBM-06-0152 for (a) chromosome12 and (b) chromosome 7.
52
-bc
m
EGFR
929kb~ 80 copies
da
-f
-h
-ie
-jg
-lk
f3
-d8-10
-f1-h6-d7-8
h2
h4-d5
a2
d1
-d3
k2
CDK4MDM2
MDM21,269kb
~ 40 copies
h1 2 3 4 5 6
8040
0
a.
b. c.
a b c d e f g h i j k l
m
1 2 1 2 3 4 5 6 7 8 9 1 2 3 1 2 3 4 5 6 1 2 3
~80~40
~80
10
~120
chr12
chr7
copy number
copy number
Figure 4.6: (a) Diagram of the two circular solutions to breakpoint graph for GBM-06-0152 that accounts for all observed, highly-supported breakpoints. (b) Circular assem-blies of the two circular solutions, possibly indicating the presence of two independentdouble minute chromosomes. The larger circular solution has two copies of MDM2 andone copy of CDK4, while the smaller solution has EGFR. (c) A zoomed-in browser shotof the amplified segment labeled h, showing the distinct changes in copy number asdescribed in (a).
53
two times, and one segment was traversed three times. The complex rearrangements fea-
tured in the MDM2-CDK4-DM are especially noteworthy in that the multiple-traversed
segments feature genomic regions connected in ways that are found only in the pu-
tative DM. This suggests that the solution shown here may not represent the initial
configuration of this sample’s DM, but rather represents the final result of an evolu-
tionary process beginning with the initial (unobserved) DM featuring, at minimum, a
single copy of each amplified segment observed in the MDM2-CDK4-DM. Subsequent
rearrangements of the initial DM then led to these initial segments being amplified
or deleted, where the more “successful” DM configurations, such as the configuration
observed with two copies of the MDM2 oncogene, would rapidly sweep through the
emerging tumor population. Since the expected increase in copy number was also ob-
served for the multiply traversed segments, the solution presented here likely represents
the dominant DM configuration in the GBM-06-0152 tumor, as shown in Figure 4.6(c).
EGFR Amplification in GBM-06-0145
The GBM tumor sample GBM-06-0145 presents both a simple and complicated
story. As shown in Figure 4.7, the amplicon containing EGFR has what appears to
be uniformly high copy number and we find an extremely well-supported breakpoint
(breakpoint support = 2,833 reads) bounding this amplicon. Upon first glance, this
data would suggest that this amplicon exists as a simple circular DM as shown in the
previous two tumors, or in a tandem array of duplicates.
However, the browser shot shows 6 additional somatic breakpoints with read
54
support > 100 and the copy number data appears to subtly change near the ends of each
breakpoint. A diagram of this same data is shown in Figure 4.8. The solution to the
breakpoint graph shown in Figure 4.8(a) shows the possibility of three distinct paths that
incorporate all breakpoints and explain the changes in copy number. All paths utilize the
bounding tandem-duplication-like breakpoint, where path A is characterized by a simple
direct path through the amplicon, path B is a similarly simple path that features a small
deletion, and the final complicated path C that incorporates the remaining breakpoints.
According to the baseline copy number, path A appears to be the dominant path, found
in a ratio of 7 to every 9 copies of the amplicon (77%).
Figure 4.8(b) shows how the structure of the EGFR gene is affected in each
path. Since path A features no internal breakpoints, EGFR is intact. On the other
hand, paths B and C alter the structure of EGFR. Path C deletes the first exon of
EGFR, likely leading to a non-functional version of EGFR. More interesting is path
B, whose internal deletion specifically removes exons 2-7 of EGFR, creating a highly
oncogenic, constitutively active form of EGFR called EGFRvIII that is found expressed
in multiple tumor types [55]. This is interesting since it suggests two scenarios: (1)
EGFRvIII emerges after wildtype EGFR is significantly amplified or (2) EGFRvIII is
created early but cells with more copies of wildtype EGFR have a selective advantage
in the tumor population. The former scenario seems most plausible, as the increased
number of copies subsequently improves the chance that the EGFRvIII mutant will
arise. Regardless of which scenario is correct, the ratio of EGFRvIII to wildtype EGFR
suggests that high copy number of oncogenic EGFRvIII may not be necessary to provide
55
significant advantage to the growing tumor cell. It would be useful to compare the
relative expression of EGFRvIII to EGFR transcripts using RNA-Seq data, if it becomes
available for this sample, to determine if EGFRvIII is preferentially expressed over the
multiple copies of wildtype EGFR present in this genome.
Amplified germline breakpoints
Figure 4.7: Browser shot of the intra-chromosomal, highly-supported breakpoints andcopy number observed for TCGA sample GBM-06-0145 for chromosome 7.
4.3.3 FISH confirmation of double minutes
For two tumors, GBM-06-0648 and GBM-06-0145, Fluorescence in situ hy-
bridization (FISH) was performed to determine the physical locations of the amplified
regions in these tumors. Depending on the pattern of the amplified probes presented
by the FISH images, extreme amplification via DMs and/or HSRs can be determined.
Figure 4.9 show the FISH results for GBM-06-0648 and GBM-06-0145, using
probes targeting sequences in the amplified and nearby, non-amplified regions of the re-
56
EGFR
Exons 2-7
a b c d e f g h i
a b c d e f g h i
EGFR
EGFRvIII
~7x~1x
~1x
a b c d f g h i
EGFRΔexon1a c d f gh ie h
Path APath B
Path C
(77%) Path A:
(11%) Path B:
(11%) Path C:
chr7
~90
~80
~100
copy number
a.
b.
2833
462
526 721
301 242
Figure 4.8: (a) Diagram of the three paths that together form a potential solution ofthe breakpoint graph for sample GBM-06-0145 that accounts for all observed, highly-supported breakpoints. The location of the gene EGFR is noted. Note that all paths,A, B and C, can create circular solutions by using the encompassing tandem duplicationshown in red to connect the end of each path to the beginning. The tandem duplicationmay also connect path A to path B, path C to path B, etc. (b) The effects of eachpath on the EGFR gene, showing that path B creates an oncogenic form of EGFR,EGFRvIII, through specific deletion of exons 2-7.
57
spective tumors. For GBM-06-0648, the amplified probe (colored red) targeted MDM2,
while the non-amplified probe (colored green) targeted the centromere of chr12. These
two probes would show the locations of these two regions in the genome, which for
normal cells would show both probes to be in single copy per chromosome and linearly
paired with each other (assuming cells are in interphase). From these results, the tech-
nicians at the cytogenics laboratory of the Memorial Sloan-Kettering Cancer Center
(MSKCC) who produced this FISH data, confirmed the presence of the MDM2 double
minute for GBM-06-0648, since the FISH image for this tumor exhibits the character-
istically distributed, punctuate pattern of an amplification by double minute.
The story is less clear for the EGFR amplification in GBM-06-0145. For this
FISH experiment, the probes targeted the amplified EGFR (red) and the centromere
of chr7 (green). Similar to sample GBM-06-0648, the amplified probe is not found in
single copy or found proximate to the non-amplified probe on chromosome 7, indicating
that the extra copies of EGFR were not caused by simple tandem duplication at the
reference location of EGFR. The technicians at MSKCC believe its pattern is primarily
HSR-like, as most cells exhibit dense clusters of the amplified probe, but it is clear
from this slide that there are cells that also display the punctuate pattern of a DM.
It is known that HSRs often spawn DM, and vice versa, so this may indicate that the
amplified state of the EGFR is in flux, alternating being present in HSR and DM form
from cell to cell [84]. The fluctuating status of this amplicon may also explain why the
breakpoint graph for GBM-06-0145, as shown in Figure 4.8, was more complicated than
the other examples.
58
b.
a.
Images courtesy of Dr. Suresh Jhanwar and Melissa A. Wynne, Cytogenetics Laboratory, Memorial Sloan-Kettering Cancer Center. FFPE blocks matching TCGA samples 06-0145 and 06-0648 were provided by Dr. Tom Mikkelsen, Hermelin Brain Tumor Center, Henry Ford Health System
Figure 4.9: Fluorescent in situ hybridization (FISH) images for samples used to deter-mine the physical state of the amplified regions. (a) FISH image for GBM-06-0648 usingprobes for MDM2 (red) and the centromere of chromosome 12 (green). Note that theimproved focus of the image on the right displays the punctuate pattern in red, char-acteristic of a DM, confirming that MDM2 is amplified by DM, as predicted. (b) FISHimage for GBM-06-0145 using probes for EGFR (red) and centromere of chromosome7 (green). Note that the red probes in most cells are clustered together, as opposed tothe pattern seen in (a), indicative of an HSR. However, there are cells in (b), such asthe bottom-left cell in the right image of (b), that display a DM-like pattern similar tothat observed in (a). This suggests that, while the dominant pattern indicates EGFRamplification via HSR, the EGFR amplicon may also take the form of double minutechromosomes.
59
4.3.4 Evidence of tumor-specific reads in blood
Summarized in Figure 4.10 is the evidence found of tumor-specific breakpoints
in the matched-normal sequencing data of the three GBM tumors discussed above. In
both GBM-06-0152 and GBM-06-0648, we find significant evidence for the majority
of DM-specific breakpoints in the matched-normal (blood) from the same patients, an
example of which is shown in Figure 4.11. In GBM-06-0152, 11 of the 20 breakpoints
have at least one supporting read in blood, while 9 breakpoints have no read support
in the blood. While all DM-specific breakpoints for sample GBM-06-0648 have at least
one supporting read in the blood, 13 of its 16 breakpoints are found to have at least 5
supporting blood reads.
Given the propensity of DMs to be lost after successive rounds of mitosis, it
seems unlikely that the DM was present in the patient’s germline and selectively retained
only in the cells that eventually became oncogenic. A more plausible interpretation is
that this oncogenic DM is instead somatic in origin, constructed and amplified at some
point prior to or during tumorigenesis. The presence of tumor discordant reads in the
blood of DM-specific breakpoints then suggests that these GBM-borne DMs cross the
blood-brain barrier to enter the patient’s bloodstream.
The third GBM tumor shown in Figure 4.10, GBM-06-0145, exhibits an ex-
treme level of EGFR amplification that could indicate the presence of an EGFR-DM.
However, FISH analysis presented above shows that amplified EGFR displayed a pri-
marily HSR-like pattern in tumor cells, with a minority of cells with the pattern of
60
DM-based amplification. Interestingly, only one breakpoint associated with the ampli-
fied region was discovered in GBM-06-0145’s matched-normal sequencing data.
The discovery of DM-specific DNA in blood begs the question of how DMs
could enter the patient’s bloodstream. One possible transport mechanism for these
oncogenic DMs is via microvesicles (or exosomes), which brain tumors are known to
produce possibly as a method for intercellular communication of mRNA transcripts
[88]. Microvesicles bud off from tumor cells with lipid bilayers intact, carrying cyto-
plasmic contents of the tumor cell such as mRNA, miRNA, and angiogenic proteins,
which can then be observed in the patient’s blood [76]. More recently, studies have dis-
covered isolated microvesicles containing single-stranded DNA (ssDNA) with amplified
oncogenic sequences, in particular c-Myc [5]. Tumor nucleic acids leaked into the blood
in this manner or via macrophages that engulf necrotic or apoptotic cells have been
proposed as possible disease biomarkers in several cancers including glioma [71, 44].
If microvesicles are the primary mechanism that carry the contents of the tumor cy-
toplasm into the blood, this may help explain why little evidence of GBM-06-0145’s
primarily HSR-amplified region was detected in blood. Since HSR-amplified regions are
integrated in the tumor genome and thus confined to the nucleus, they are far less likely
to be included in a budding microvesicle than the cytoplasmic DMs.
61
48
121620
5001000150020002500
h b w p s c r q m o l n t f k d
246810
10002000300040005000
a d f h i e j g l k f3 d8-9 f1 h6 d7-8 h2 h4 d5 a2 d1 d3 k2 c m b
a.
b.
10002000300040005000
2468
10
5001000150020002500
48
121620
MDM2-CDK4 DM EGFR DM
segment
segment
MDM2 DM
c.
2468
10
5001000150020002500
a-i d-f a-h h-c f-d d-g
5001000150020002500
connection
2468
10
EGFR DM
Figure 4.10: Bar graphs showing breakpoint read support for tumor (black bars) andblood (red bars) for sample: (a) GBM-06-0152, (b) GBM-06-0648, and (c) GBM-06-0145.
62
a.
b.
Tumor-specific discordancies in blood
Deletion-type Somatic Breakpoint
Figure 4.11: Browser shot of raw sequencing reads for GBM-06-0648 surrounding theleftmost position of a highly-supported deletion-type breakpoint that is part of thissample’s putative DM. (a) The position of the deletion-type breakpoint and an excerptof the tumor sequencing reads that support the breakpoint and the associated amplicon.(b) The sequencing reads for the matched-normal (blood) sample for GBM-06-0648showing a set of 9 sequencing reads that support the breakpoint shown in (a).
63
4.4 Discussion
The integration of relative copy number with structural variants as described
here has demonstrated how such data can be used to improve our understanding of the
genomic topology of the cancer genome. We show that, in the case of highly amplified
regions of the tumor, we are able to completely explain both the observed copy number
and highly supported breakpoints by solving a simple breakpoint graph, which describes
the order and orientation of the highly amplified segments in the tumor genome. We can
obtain local genome assemblies that are accurate down to the base-level and precisely
determine their genetic content, an impressive achievement considering this is done
using only a pair of short-read sequencing datasets.
In two of the GBM samples discussed here, the optimal solutions to the break-
point graphs of amplified segments are circular. These circular solutions suggest that
the observed amplified regions are DMs. The presence of oncogenes on each DM and
their highly amplified state indicate that the DMs have strong oncogenic potential, con-
fer a selective advantage to the tumor cell, and their formation was likely a key event
in the tumorigenesis of these GBM tumors. For the GBM tumor with FISH data, the
predicted DM was confirmed to exist. A third sample exhibiting extreme amplification
of EGFR yielded a more complicated breakpoint solution, potentially supported by the
fluctuating state between HSR- and DM-like patterns observed in its FISH results.
Most interesting, nearly every breakpoint specific to the DMs of two GBMs
also has detectable read support in the blood of the patient. In fact, the infiltration of
64
GBM-borne DMs in the blood for sample GBM-06-0648 is observed at such a high level
that DM-specific breakpoints could be reliably inferred using only the whole genome
sequencing data derived from the patient’s blood. This result suggests some intriguing
clinical applications, which will be described further in Section 6.2.
65
Chapter 5
Inferring clonality and tumor
evolution
A tumor growth is a population of cancer cells. This population may homoge-
nous, where all tumor cells share similar characteristics. Such tumors are defined here as
“monoclonal” since all tumor cells feature the same variants (copy number aberrations,
structural variants, mutations) as each other, all clones of the same progenitor tumor
cell. This progenitor tumor cell may be the first cancerous cell that initiated the tumor,
or may be a subsequent tumor cell that gained an advantageous mutation that aided a
complete sweep of the tumor population.
“Polyclonal” tumor growths are defined as tumors composed of at least two
distinct clonal populations of tumor cells. In “polyclonal” tumors, each clonal popu-
lations arose from a clone, but each clone differs from the other by some observable
66
alteration. The multiple clonal populations may be significantly different from each, or,
as is more often the case, the clonal populations are related, sharing a set of variants
that are found in all or a large subset of tumor cells. For the purposes of this analysis,
we will define a polyclonal tumor to consist of multiple “major clones,” where “major
clone” represents a “detectable” clone, typically representing ∼10% of the tumor popu-
lation. This distinction is necessary to remind us to remain cognizant of the fact that a
growing tumor population is not static and is likely to consist of numerous minor clones
that are undetectable with any given method.
In addition to the above definition of tumor clonality, we can further classify
individual mutations as either “clonal” or “subclonal.” When the dominant clones of a
particular tumor is found, “clonal” variants are those shared by all tumor cells of any
or all dominant clones, in other words these variants achieved full penetrance in the
entire population or polyclonal subpopulation of cells. “Subclonal” variants are those
that exist in only a small proportion of the cells belonging to a clonal population.
Let’s apply the above definitions to an example of a tumor’s evolution, shown
in Figure 5.1. Suppose an initial cell acquired a nonsense mutation in a key tumor
suppressor (M1) and amplified an oncogene (A1) that supported the initial growth of
a tumor. Early on in this tumor’s development, another tumor suppressor was deleted
(D1) that caused the tumor cell to grow even more rapidly, enabling cells with this
deletion to rapidly overtake the entire tumor population. Soon after acquiring deletion
D1, a cell also acquires a set of neutral mutations (M2, M3), amplifications (A2, A3),
deletions (D2, D3). Since these variants occurred early during the clonal expansion of
67
this tumor cell variant, but do not provide any selective advantage, the population of
tumor cells are split into two “major clones,” where 25% of tumor cells have the neutral
variants (M2, M3, A2, A3, D2, and D3) and 75% of tumor cells do not. Much further
during this tumor’s development, additional mutations (M4, M5) appear on one of the
two major clones, but do not have a chance to spread through the population prior to
the patient’s death and/or tissue biopsy.
In the above example, we would define this tumor population as polyclonal,
with its two major clones defined such that: clone (1) has variants M1, A1, and D1,
and clone (2) shares the variants of clone (1), but in addition has variants M2, M3, A2,
A3, D2, and D3. The clonal mixture is determined as 75% clone (1) and 25% clone (2).
Mutations M1, M2, and M3 would all be classified as “clonal” since they all achieved
full penetrance in their respective clones, whereas M4 and M5 would be classified as
“subclonal” mutations.
This chapter presents various methods to extract and synthesize data to recon-
struct the clonal evolution of a tumor from whole genome sequencing data of a single
tumor biopsy. In a mix of automated and manual processes, these methods form a
framework to determine the tumor’s clonality, the number and proportion of all major
clones, and the variants that distinguish them. Furthermore, a method to phase muta-
tions to parental alleles to time their emergence within the population is described. In
addition, the methods will provide an accurate estimate of the amount of contaminating
normal tissue was present in the tumor’s biopsy.
68
Germline
+ M1 mutation
+ A1 amplified oncogene
+ D1 deletion of tumor suppressorInitial Tumor Cell
Clone A (~75%)
+ M2, M3 passengers+ A2, A3 passengers
+ D2, D3 passengers
Clone B (~25%)
Undetectable “Minor Clone”
Normal Tissue
Tumor Biopsy
+ M4, M5
Figure 5.1: A toy example of the evolution of a tumor starting from the germline, toan initial tumor cell, to a population of major and minor clones that are sampled bythe tumor biopsy. To avoid complicating this example, mutation and copy numberpassenger events that occurred prior to the emergence of the initial tumor cell were notincluded.
69
5.1 Copy number alterations, allele fraction, and the al-
lelic state diagram
To discover and describe the major clones of a population, relative copy number
and allele fraction estimates made by BamBam of copy number alterations are utilized.
Other groups have studied tumor clonality using SNP chips or single point variants
sequenced at extremely high coverage (> 150x for whole genomes or> 1000x for exomes)
[15, 59, 92]. Unfortunately, the precision of allele fraction estimates is poor for point
mutations using the average coverage (30 − 50x) whole genomes sequenced by TCGA,
which is a key measure in determining clonality. By aggregating over large genomic
regions, the error in these estimates is drastically reduced, enabling distinct clones to
be resolved. However, this does limit the ability of this method. It can resolve major
clones only based on copy number alterations. Clones distinguished solely by point
mutations are not discoverable by this method, although this limitation is mitigated by
the mutation clonality analysis described below in Section 5.4.
Underlying the method to determine both clonality and estimate normal con-
tamination is the “allelic state diagram” (ASD), which is defined here. The ASD de-
scribes the positions of clonal allele-specific copy number variants using both relative
copy number and allele fraction of copy number alterations, demonstrating the rela-
tionship between copy number and allele fraction for all allelic states. The positions of
70
clonal allelic states in the ASD are determined by the following:
CN(tmaj , tmin, nmaj , nmin, α) =(1− α)(tmaj + tmin) + α(nmaj + nmin)
nmaj + nmin(5.1)
AF (tmaj , tmin, nmaj , nmin, α) =(1− α)tmaj + αnmaj
(1− α)(tmaj + tmin) + α(nmaj + nmin)(5.2)
where CN is the relative tumor copy number compared to a matched-normal, AF is
allele fraction in the tumor, α is the fraction of normal contamination in the tumor
sample, and tmaj , tmin, nmaj & nmin are the majority and minority allelic states in the
tumor and normal, respectively. Since individual genomes can only have discrete allelic
states, such that they have 0, 1, 2, or more copies of a given chromosomal segment,
the possible values for tmaj and tmin are constrained to the set of positive integers,
ti ∈ (0, 1, 2, ..., n). Furthermore, the majority and minority allelic states for the normal
are set to one, ni = 1, which is true for all of the autosomes in a normal human genome.
The sex chromosomes, X and Y, are currently ignored in the ASD. Note that since
the above math necessarily requires two alleles, only heterozygous sites in the matched-
normal genome are considered for the ASD.
The forthcoming figures all share the same set of allelic states: normal copy
number, single-copy amplification, single-copy / hemizygous deletion, homozygous dele-
tion, copy-neutral LOH (CN-LOH), and amplification of both parental alleles. Fig-
ure 5.2 shows simulated copy number and allele fraction data for these allelic states
with no normal contamination, demonstrating how the ASD can be used to determine
the allelic state of each cluster of points. Each vertex in the ASD’s grid is labeled
with its tumor allelic state, (tmaj , tmin), and the position is determined by the equa-
71
tions above. Figure 5.3 demonstrates how the location of allelic states are affected by
increasing amounts of normal contamination, α. As normal contamination increases,
the allelic state positions grow closer together, reducing the ability to resolve different
allelic states.
(1,1)
(2,1)
(3,1)
(4,1)
(5,1)
(6,1)
(2,2)
(3,2)
(4,2)
(5,2)
(3,3)
(4,3)
(2,0)
(1,0)
(3,0)
(4,0)
(5,0)
(6,0)
(7,0)
(0,0)
Normal
Single-Copy Amp.
Multi-Copy, Biallelic Amp.
Homozygous Del.
Hemizygous Del.
CN-LOH
Majority Allele Fraction
Rela
tive
Cop
y N
umbe
r
(# Majority Copies, # Minority Copies)
Figure 5.2: An example of an allelic state diagram (ASD) of simulated data for a mon-oclonal tumor sample with zero normal contamination, α = 0. Chromosomal regionsexhibiting different copy number alterations are plotted in different colors. This simu-lated tumor genome exhibits 6 allelic states: normal (black), single-copy amplification(yellow), hemizygous deletion (red), homozygous deletion (dark green), copy-neutralLOH (CN-LOH, blue), and multi-copy biallelic amplification (cyan).
The case presented in Figure 5.3 is a static snapshot of a monoclonal tumor.
72
a. b.
c. d.α = 0.9
α = 0.1
α = 0.5
α = 0.0
Normal
Single-Copy Amp.
Multi-Copy, Biallelic Amp.
Homozygous Del.
Hemizygous Del.
CN-LOH
Majority Allele Fraction Majority Allele Fraction
Majority Allele Fraction Majority Allele Fraction
Rel
ativ
e C
opy
Num
ber
Rel
ativ
e C
opy
Num
ber
Rel
ativ
e C
opy
Num
ber
Rel
ativ
e C
opy
Num
ber
Figure 5.3: A set of allelic state diagrams for the same simulated monoclonal tumorgenome shown in Figure 5.2 with different levels of normal contaminant: (a) 0%, (b)10%, (c) 50%, and (d) 90%. Notice that, as the level of normal contamination in-creases, the distance between allelic states is reduced, negatively affecting our ability todistinguish between different states.
73
However, we know that the tumor genome can be very dynamic, with gains and losses of
small and large chromosomal segments. Figure 5.4 shows some of the possible transitions
between the allelic states described in previous figures. Note that some transitions are
deemed “one-way” since they involve the irreversible loss of chromosomal segments.
For example, the transition between the normal allelic state (1,1) and the hemizygous
deletion state (1,0) is “one-way” because that deleted allele can never be restored.
However, the retained allele in this case can be amplified, permitting transitions to
the copy-neutral LOH (CN-LOH) state and beyond (2+,0). Notice that the deletions
necessary for the transition between the cyan-colored allelic states are not deemed “one-
way” because at least one copy remains of each allele remains in the genome.
Figure 5.5 displays the ASD of a tumor genome transitioning from the allelic
states presented in the previous figures to new allelic states that differ only by a single
copy loss or gain. During such a transition, the population of tumor cells will be a
mixture of tumor cells consisting of tumor cells with the original allelic states and
tumor cells with the new allelic states. For the example presented in Figure 5.5, we can
think of this “transitional” tumor as a population divided between two major clones,
A and B, where clone A is defined by the original allelic states, and clone B is defined
by the new allelic states. The mixture fractions shown on this figure, Mb, represents
the fraction of clone B within the population, such that the tumor population solely
consists of clone A when Mb = 0, and the population consists only of clone B when
Mb = 1. It is important to note that the allelic states for both clones, ti,a and ti,b, are
still constrained to the set of positive integers.
74
(1,1)
(2,1)
(3,1)
(4,1)
(5,1)
(6,1)
(2,2)
(3,2)
(4,2)
(5,2)
(3,3)
(4,3)
(2,0)
(1,0)
(3,0)
(4,0)
(5,0)
(6,0)
(7,0)
(0,0)
Majority Allele Fraction
Rel
ativ
e C
opy
Num
ber
(# Majority Copies, # Minority Copies)
One-way
One-way
One-way
Figure 5.4: Allelic state diagram of the case presented in Figure 5.2 showing some possi-ble bidirectional and unidirectional transitions between its allelic states. Note that uni-directional transitions are those that involve irreversible loss of a parental chromosome.
75
gain of non-amplified allele
single loss of amplified allele
gain of remaining allele
0%25%
50%75%
100%
0%25%
50%75%
100%
0%25%
50%75%100%
Majority Allele Fraction
Rel
ativ
e C
opy
Num
ber
(2,1)
(2,2)
(4,2)
(5,2)
(2,0)
(1,0)
Figure 5.5: Allelic state diagram of the case presented in Figure 5.2 showing how tran-sitional allelic states are created when the tumor consists of a mixture of two differentclones. In this case, there are two subclones. Clone A is defined by the original allelicstates: yellow (2,1), cyan (5,2), and red (1,0). Clone B alters these states through am-plifications and a deletion to produce the allelic states: yellow (2,2), cyan (4,2), and red(2,0). They share the same allelic states of blue (2,0), black (1,1), and green (0,0). Thepercentages denote the percentage of clone B present in the tumor population, where0% describes a monoclonal population of clone A, and 100% is a monoclonal populationof clone B.
76
From Figure 5.5, when the mixture fraction Mb is such that the tumor is a
heterogenous population of cells, Mb = 0.25, 0.5, 0.75, the allelic states do not lie on the
vertices of the ASD, but rather on the edge connecting two vertices. A tumor population
in such a state would be classified as polyclonal. Take, for example, the cluster of red
points in Figure 5.5. In this region of the genome, clone A has the allelic state of
hemizygous deletion, or (1, 0), whereas clone B has amplified clone A’s retained allele,
altering its allelic state in this region to copy-neutral LOH, or (2, 0). When Mb = 0,
the allelic state of the red points are found clustered on the ASD vertex representing
the hemizygous deletion allelic states. As M increases, meaning increasing amounts of
clone B in the population, the cluster of points progresses along the edge towards the
CN-LOH state. At Mb = 0.5, where there are equal amounts of clone A and B in the
population, the cluster of points is found precisely in the middle of the edge between
the LOH and CN-LOH allelic states.
If the tumor population consists of non-derivative clones, or clones that are
distantly related to one another such that their allelic states do not differ by single copy
gains or losses, the position of the mixture of allelic states will not lie along the edges
of the ASD, as shown in Figure 5.6. We will see later that such abnormal allelic states
can also occur when more than 2 major clones exist in a polyclonal tumor.
This demonstrates how the ASD can, at a glance, indicate the presence of
one or more major clone in a tumor sample, determine the allelic states of the major
clones, and provide a visual estimate of the proportion of each major clone in the tumor
population. Thus, the ASD can, at a minimum, be used as a powerful visual diagnostic
77
0%
25%
50%
75%
100%
0%25%
50%75%
Majority Allele Fraction
Rel
ativ
e C
opy
Num
ber
0%
25%
50%
75%
0% 25%
75%
100%
100%
75%
50%
25%0%
50%
(1,0)
(2,0)
(3,0)
(4,1)
(2,1)
(5,2)
(1,1)
(0,0)
Figure 5.6: Allelic state diagram of the case presented in Figure 5.2 showing the tran-sitional allelic states produced when allelic states are “skipped,” which can occur whenthe tumor consists of two or more unrelated, or distantly-related clones. Notice thatthe transitional states are not found on the edges connecting allelic states if the allelicstates of the two major clones differ in both majority and minority alleles.
78
tool for determining the clonality of a tumor sample.
5.2 Fitting sequencing data to the ASD
The mathematics behind the ASD, Equation 5.1 and Equation 5.2, is modeling
the ideal case where the relative copy number is 1.0 and the majority allele fraction is
0.5 for “normal” (1, 1) allelic states. However, the results produced by BamBam on real
world data do not precisely fit this idealized case. To estimate relative copy number,
BamBam calculates the relative coverage between tumor and normal. If the tumor
and normal samples are sequenced at the same coverage level, relative coverage is an
accurate measure of relative copy number. However, this will not be the case if the
tumor sample is sequenced at a much higher coverage than its matched-normal, in an
attempt to improve detection of mutations, particularly subclonal mutations, in the
tumor sample.
For example, assuming no normal contamination, if a tumor is sequenced at
twice the coverage of its matched-normal, then a region with a “normal” allelic state
will have twice as many reads in the tumor as it has in the normal. Thus, this region
has a relative coverage of 2.0 and a relative copy number of 1.0, and BamBam’s use
of relative coverage will not fit the ASD. Unfortunately, the precise coverage level of
a given sequencing dataset is unknown, as the sequencing centers merely target the
desired coverage level, but have no guarantee of achieving it. Using the raw number
of reads found in the tumor and matched-normal datasets as an estimate of overall
79
coverage level can help correct the imbalance, but is complicated by the ploidy of the
tumor sample. If a tetraploid tumor (ploidy = 4.0) and its matched-normal (ploidy =
2.0) are sequenced at the same physical coverage, the tumor will have two times the
number of raw reads than the matched-normal. So, using the ratio of their raw numbers
of reads to scale local relative coverage estimates would this tetraploid tumor to appear
to have normal copy number.
The error in the estimate of allele fraction that BamBam produces is caused
by a limitation in how the majority allele is selected in regions of allelic balance, such
as the “normal” allelic state. Ideally, the allele fraction for such regions should be
approximately 0.5, but this only occurs when both alleles have equal read depths. More
often, due to the stochastic nature of how heterozygous alleles are sampled from the
pool of genomic DNA, one of the two alleles will likely have a slightly higher read depth
than the other, causing a slight increase in the majority allele fraction that is estimated.
For example, assuming no normal contamination, a whole genome with 30x
coverage would ideally produce 15 reads of each allele at heterozygous “normal” allelic
states. However, if one allele’s read depth was shifted by just a single read, such that
allele A has read support of 16, BamBam would estimate the majority allele fraction to
be 1630 = .53, a deviation of 0.03 from the actual allele fraction. Usually averaging over
multiple positions can reduce the effect of such errors, however the error in majority
allele fraction for these balanced allelic states cannot be averaged out because majority
allele fraction, by its definition, can never dip below 0.5. Fortunately, sampling error
has a much less pronounced effect on amplified and deleted allelic states. In these cases,
80
the majority allele is readily identifiable and the sampling error can be averaged out
over multiple positions.
In order to fit BamBam results onto the idealized ASD, we must then model
and correct out the above errors from the data. The model has four parameters: normal
contamination α, allele fraction delta AFd, coverage delta COVd, and coverage scaling
factor COVs. The α parameter only affects grid layout of the ASD, as shown in Fig-
ure 5.3. The latter three parameters transform the BamBam results. The paramaters
COVd and COVs affect the y-axis shift of the copy number data and scale of copy
number data from the “normal” allelic state according to the following equation:
CNcorr(CN,COVd, COVs) = COVs(CN − COVd) + 1.0 (5.3)
where CN is the relative copy number estimate produced by BamBam and CNcorr is
the corrected copy number used to compare against the ASD. The final parameter, AFd
has its strongest effect on the the allele fraction estimates of the allelic balanced states.
It does this with the following equation:
AFcorr(AF,AFd, CNcorr, x) = AF − AFdCNcorr
(1.0− (AF −AFd)
0.5)x (5.4)
where x is set to a large integer (e.g. x = 20) to rapidly reduce the degree to which allele
fraction estimates are corrected as they diverge from balanced allelic states. This param-
eter helps reduce the error in allele fraction estimates introduced when the “majority”
allele achieved a higher read support from unbalanced sampling of the heterozygous
site, such as the normal allelic state, (1, 1), and any amplified state where both alleles
are amplified equally, (N,N). It is very important to not appreciably alter the allele
81
fraction estimates at deleted states, as they are the determining factor for estimating
normal contamination.
The optimal values for these four parameters are discovered using gradient
steepest descent search, optimizing the RMSD of the corrected copy number and allele
fraction estimates, CNcorr and AFcorr, to the ASD defined by the normal contamination
parameter, α. The search begins with a set of initial values for each parameter, and a
set of increments for each parameter, COV id , COV
is , AF
id and αi. For each parameter,
p, and parameter increment, pi, the RMSD from the ASD is calculated for p, p + pi
and p− pi. The parameter value that yields the greatest reduction in RMSD among all
four parameters is chosen as the new current value for that parameter, and the cycle
repeats. If no reduction in RMSD is possible with the current parameter increments, the
increments are divided by half and the search resumes. Once three rounds of divisions
have occurred, the search is concluded and the best fit parameters are reported.
Since gradient descent can often get stuck in a local minimum, gradient search
is performed with a number of different initial parameters until a consistent set of fit
parameters is found.
5.3 Modeling the clonal mixture of a tumor sample
The ASD can be used to determine a set of allelic state “landmarks” that
help define the number of clones and their proportions within the tumor population,
Li = (CNcorr,i, AFcorr,i). The landmarks used in this analysis will be defined by the
82
large clusters of points on the ASD, as they indicate major portions of the tumor that
have undergone a copy number change in a significant fraction of the overall tumor
population. See Figure 5.7(b) for the landmark allelic states used to analyze GBM-06-
0185. For each landmark on the ASD, all plausible clonal mixtures that would result
in its observed copy number and allele fraction are considered, then the optimal clonal
mixture is chosen such that it can account for all ASD landmarks most parsimoniously.
As observed in Figure 5.5, for monoclonal tumor populations, we expect the
landmarks to all lie on ASD vertices. In polyclonal tumors consisting of two major
clones, where clone B inherits all of clone A’s allelic states and has additional allelic
states distinct from clone A, we will find landmarks on both the vertices and edges of
the ASD. The landmarks that lie on vertices are those that represent the allelic states
shared by both clone A & B, while the landmarks on ASD edges represent the mixture
of the different allelic states. The position along this connecting edge determines the
proportions of clones A and B in the mixture. If multiple landmarks are found on
edges and not on vertices, then the variety of positions along their respective edges will
determine the number of clones.
For example, if all landmarks are found halfway between two allelic states,
that is most simply explained by two major clones in equivalent proportion within the
population. If, however, one landmark is located at the halfway mark and another is
found 25% along the way towards an allelic state, there must be more than two clones
in the population. One simple explanation for this is that there are three clones, A, B,
& C, where A makes up 50% of the tumor population and clones B & C make up 25%
83
each. Suppose clones B & C both exhibit a single-copy allelic state change from clone
A, explaining the halfway landmark. The 25% landmark is then explained if, in that
chromosomal segment, clone B (or C) experienced a single-copy allelic state change not
found in clones A and C (or B).
Thus, the problem at hand is determining the least number of major clones
that explain n observed landmarks:
Lobs ∈ (Lobs0 , Lobs1 , ..., Lobsn ), (5.5)
where Lobsi = (CNobsi , AF obsi ). We assume a mixture of m clones, each with k integral
majority and minority allelic states Ci = [(t0maj,i, t0min,i), (t
1maj,i, t
1min,i), ..., (t
kmaj,i, t
kmin,i)],
and mixture proportions, Mi, such that∑Mi = 1.0 − α. The relative copy number
and allele fraction of each landmark, Lmixi , is a linear combination of the allelic states
indexed by i across the clonal mixture:
CNmixi =
2α+∑m
k Mk(tkmaj,i + tkmin,i)
2(5.6)
AFmixi =α+
∑mk Mk(t
kmaj,i)
2α+∑m
k Mk(tkmaj,i + tkmin,i)
(5.7)
where the normal allelic states for all clones are assumed to be nkmaj,i = nkmin,i = 1.
The optimal solution is the one that most closely approximates the observed landmarks
with a simplest mixture of major clones, or optimizing the objective function:
O(Lobs, Lmix) =1
n
n∑i
(CNobsi − CNmix
i )2 + (AF obsi −AFmixi )2 +mx (5.8)
which is the squared distance from the observed data plus a penalty, m, equal to the
number of clones in the population, that is controlled by the strength parameter x. This
84
is a complicated objective minimization problem that due to the inclusion of integral
allelic states. Shortcuts based on modeling the evolutionary process of the tumor’s
development are possible.
The optimization of Equation 5.8 is performed after finding the best fit param-
eters for α,AFd, COVd and COVs as described in Section 5.2. It begins by identifying
all “shared” landmark allelic states, which every clone in the mixture must exhibit. If
we assume a tumor is step-wise evolving, these shared allelic states represent the “root”
of the tumors evolutionary tree. If there are no landmarks on ASD edges, the procedure
is complete and the tumor population is classified as monoclonal.
If landmarks exist along ASD connecting edges, between two bounding allelic
states, then additional clones are necessary. The procedure adds one additional “daugh-
ter” clone to the mixture, which inherits all of the shared allelic states and gains an
allelic state and mixture proportion necessary to explain the edge-bound landmark. If
more than one edge-bound landmark can be explained with the same mixture propor-
tion, then those new allelic states are added to the new clone. This process is repeated
until all non-vertex landmarks are explained by the clonal mixture, wherein each addi-
tional “daughter” clone can derive from any current clone in the mixture that bounds
one side of an unexplained landmark. Once all landmarks can be reasonably explained,
the clones’ allelic states and mixture proportions are reported.
Note from Equation 5.6 and Equation 5.7 that the combination of allelic states
that uniquely determine each landmark’s position on the ASD can also determines the
phased set allelic states for all positions in the genome that correspond to the landmark.
85
This can only work when the mixture proportions are unique for each clone, i.e. the
major clones must unevenly split the tumor population. In such cases, this enables
whole genome, allele-specific copy number profiles to be inferred for each clone in the
tumor population.
5.4 Linking mutations to clone-specific allelic states
To achieve an even greater understanding of the evolution of a tumor, we should
not constrain ourselves to only looking at copy number changes. By integrating somatic
mutations into this framework, we can attempt to determine when a mutation arose
during the tumor’s development. To do this, we will first attempt to link mutations
directly to the majority or minority allele in the encompassing chromosomal region on
the ASD and then use its mutational allele fraction to determine whether the mutation
occurred prior to the change in allelic state, soon after the allelic state change, or much
later.
5.4.1 Via direct phasing
For every mutation discovered by BamBam, we will find all nearby germline
heterozygous variants and look for paired reads that physically connect, or “phase,” the
mutation allele to a specific germline allele. “Nearby” is defined as being separated by
no more than double the insert size of the paired read library, typically ∼1,000bp for
these whole genome libraries, as that is well outside the expected distance that would
86
separate two paired reads.
All read pairs that overlap the positions of both the mutation and the germline
variant are collected and the number of times the mutation is phased to either of the
two germline variant alleles is recorded. If the mutation is found linked to the same
germline variant allele more than once, and is not found also phased to the other allele
of that germline variant, it is considered to be directly phased to that germline variant
allele. Phasing can be made either within a single read if the mutation and germline
variant are separated by less than a read’s length, or can occur across mates of a read
pair. Mutations can also be phased to multiple germline variant positions.
For every mutation that can be directly phased to a germline variant, the
germline variant’s allele fraction is used to determine if the mutation is phased to the
majority or minority allele. If the germline variant’s allele fraction is determined to be
greater than or equal to 0.5, then the mutation is deemed “majority-phased,” otherwise
it is phased to the minority allele, or “minority-phased.” Note that in the cases when
the two allelic states are equal, such as normal (1, 1) or bi-allelic, balanced amplifica-
tions (2, 2), the mutation’s assignment to “majority” or “minority” allele depends on
whichever allele was sampled slightly deeper in the sequencing data. Thus, classifying
mutations as “majority-phased” or “minority-phased” in such cases is not as meaningful.
5.4.2 Via amplified allele fraction
When direct phasing cannot be made, our ability to determine which allele the
mutation is linked to is severely limited. However, when mutations are found within an
87
amplified chromosomal segment, we can use the mutation’s allele fraction to determine
to which allele the mutation may be linked. When the mutation’s allele fraction is
approximately equal to the majority allele fraction, this can only have occurred if the
mutation was present in the amplified allele prior to the amplification. If the mutation
were instead on the un-amplified allele, the mutation’s allele fraction would necessarily
be much lower.
However, low mutation allele fractions do not necessarily indicate that the allele
is not “majority-phased,” since mutations can occur post-amplification. For example, if
a region was amplified by a single copy, allelic state (2, 1), a post-amplification mutation
could be, at most, present on one copy of the majority allele, with a maximal allele
fraction of 12+1 = 0.33, compared to the expected allele fraction of a pre-amplification
mutation, 22+1 = 0.67.
Thus, we are limited to linking un-phased mutations to amplified segments
when the mutation occurs prior to the amplification. This can be still be useful, as we
expect oncogenic mutations to occur early in the tumor’s development, as they are likely
to drive tumor growth. If multiple copies of these oncogenic mutations are selectively
advantageous for the tumor cell, then we’d expect the requisite increase in mutation
copy number and allele fraction to enable us to use this method.
5.4.3 Comparing allele fractions to infer mutation timing
After assigning mutations to the majority or minority alleles, we compare the
allele fraction of the mutation to the allele fraction of the majority or minority allele
88
fraction of the chromosomal segment that encompasses the germline variant allele. We
use the allele fraction of the chromosomal segment instead of the germline variant allele
because the estimate of the chromosomal segment’s allele fraction is more accurate due
to averaging over all germline heterozygous positions within the segment.
To accurately compare a mutation’s allele fraction to the majority or minority
allele fraction of heterozygous positions, we must add in some “normal” contamination
to the mutated allele. Recall from Equation 5.2 that majority allele fraction, AF ,
features normal contamination in both its numerator and denominator. This is due
to the fact that the positions considered in these equations are heterozygous in the
normal, and thus you expect to get normal contamination from both alleles. However,
for a somatic mutation, there is no normal contamination of the mutant allele, as the
mutation does not exist in the normal:
MAF =(1− α)mtmaj
(1− α)(mtmaj + (1−m)tmaj + tmin) + α(nmaj + nmin)(5.9)
where MAF is mutant allele fraction, m is the fraction of copies of the tumor allele
tmaj that are mutated, and tmaj , tmin, nmaj & nmin represent the same homozygous
allele. To fairly compare MAF to allele fractions estimated at heterozygous positions,
we must add the following correction:
MAFc = MAF +αnmaj
(1− α)(tmaj + tmin) + α(nmaj + nmin)
=(1− α)mtmaj + αnmaj
(1− α)(tmaj + tmin) + α(nmaj + nmin)(5.10)
where MAFc is the corrected mutant allele fraction. Note that while m is allowed to be
any fraction less than or equal to zero in the above equations, there are some values of
89
m that have special meaning. If m = 1, then all of the tmaj alleles were mutated, and
in the cases where tmaj represents an amplified allele, when m = 1 the mutation must
have occurred prior to the amplification. When m = 1tmaj
, where tmaj represents the
number of copies of the amplified allele, then we know the mutation must have occurred
soon after the amplification, since it exists on a single copy of the amplified allele but
is found in this state in the majority of tumor cells. If, however, m << 1tmaj
, then the
mutation must have occurred after the amplification, likely very late during the tumor’s
growth, as its very low allele fraction indicates it’s only found in a small fraction of
tumor cells.
If the mutation is phased to the minority allele, tmin, we expect to find the
maximum mutation fraction to be m = tmintmaj
as that indicates that all copies of the mi-
nority allele were converted. So, when the minority allele state exists in single copy and
all copies of it were mutated, m = 1tmaj
, precisely the same mutant fraction one would
calculate for a “majority-phased” mutation present at single copy. Thus, only with the
direct phasing can you distinguish between an early “minority-phased” mutation and
mutations that occur post-amplification.
5.5 Results
We will test the above methods on whole genome sequencing data from two
tumor types studied by TCGA: glioblastoma multiforme (GBM) and lung squamous
cell carcinoma (LUSC). These two types of tumors were chosen as they demonstrate
90
Table 5.1: Clonality of 12 GBMs.
Sample Clonality Normal Cont. (α) # Major Clones
GBM-06-0145 monoclonal 22.5% 1GBM-06-0155 monoclonal 24.5% 1GBM-06-0877 monoclonal 29.8% 1GBM-06-0648 polyclonal 12.5% 2GBM-06-0152 polyclonal 24.1% 3GBM-06-1086 polyclonal 7.5% 4GBM-06-0185 polyclonal 14.6% 4GBM-14-1454 polyclonal 5.9% > 1GBM-14-0786 polyclonal 13.9% > 1GBM-06-0188 polyclonal 20.6% > 1GBM-06-0214 polyclonal 33.6% > 1GBM-26-1438 polyclonal 50.0% > 1
the full range of clonality, from monoclonal to highly polyclonal, possible in different
cancers.
5.5.1 Clonality of glioblastoma multiforme tumors
12 whole genome GBM samples were processed with the above methods to
determine the level of normal contamination and the clonality present in each tumor
biopsy. The relative coverage and allele fraction produced by BamBam for the other 5
whole genome GBM samples discussed in previous sections possessed too much variabil-
ity to be analyzed by these methods. The results of the clonality analysis are summarized
in Table 5.1. Surprisingly, only 3 GBM tumor samples were found to be monoclonal,
while the other 9 samples consisted of at least two major clones. For 7 GBM tumors, the
precise mixture of clones were determined, while the remaining 5 tumor were inspected
visually to determine their clonality.
91
Results of two tumors, GBM-06-0145 and GBM-06-0185, are shown in Fig-
ure 5.7. The relative coverage and allele fraction data of these two samples were trans-
formed using the best fit paramaters as described above, demonstrating close fit to the
ASD with estimated normal contamination levels of 21.5% and 14.6%, respectively. By
inspecting where the data cluster, whether on vertices or edges, we can visually deter-
mine the clonality of these tumors. Since all of GBM-06-0145’s data cluster around
ASD vertices, it is likely this tumor is monoclonal. On the other hand, GBM-06-0185
is clearly polyclonal, since the several large clusters along the ASD edges indicate the
presence of at least two major clones in this tumor. In fact, since the edge-bound clus-
ters are found at different positions along their edges (e.g. the blue clusters are at the
halfway mark, while the two purple clusters are approximately .75 and .80 of the way
towards the single-copy deletion state, respectively), this can only occur from a mixture
of at least three major clones.
To determine precisely the number of clones in these samples, we used the
methods described above to determine the number of clones and their allelic states.
For each inferred clonal mixture, we computationally determine relative copy number
and allele fraction for every position in the genome given the derived clonal mixture
and compare it against the results produced by BamBam. This provides a metric to
determine how well the clonal mixture models the observed data.
As shown in Figure 5.8, we find a single clone for GBM-06-0145, as expected.
The computationally-derived relative copy number and allele fraction data shows a very
good fit to the observed data. We find a total of four major clones for GBM-06-0185,
92
GBM-06-0185
GBM-06-0145a.
b.
α = 0.215
α = 0.14625
Majority Allele Fraction
Majority Allele Fraction
Rel
ativ
e C
opy
Num
ber
Rel
ativ
e C
opy
Num
ber
Figure 5.7: Allelic state diagrams for two GBM tumors: (a) GBM-06-0145 and (b)GBM-06-0185. The fitted parameters found normal contamination at 21.5% and 14.6%,respectively. Notice that in (a) there are only clonal allelic states and no evidence oftransitional allelic states, indicating that GBM-06-0145 is a monoclonal tumor. How-ever, in (b), there are both clonal states and multiple transitional allelic states. Sincethe blue and purple transitional allelic states feature three different mixture percentages,this polyclonal tumor must consist of at least three subclones. The black X’s plotted in(b) represent “landmark” allelic states that will be used to determine the clonal mixtureof GBM-06-0185.
93
whose clonal allelic profiles (CAPs) are presented in Figure 5.9. There are two important
things to note from the four clones presented here. Firstly, as described before, the fact
that each clone’s mixture proportions is different from all others helps us to phase the
allelic states across the whole genome into allele-specific copy number profiles for each
clone.
Secondly, all clones appear to have derived from clone A. Each derivative
clone shares all of the events found in clone A, suggesting that clone A is the progenitor
of clones B, C, and D. It’s unclear if this set of clones evolved linearly, in a step-
wise progression, or if clone B and clones C & D represent different lineages. These
latter three clones differ by the deletions in chr6q, where clone B features a set of focal
deletions while clones C & D have lost all of chr6q. These are not mutually exclusive
events, so it’s possible that clone C had derived from B, inheriting its focal deletions and
subsequently deleting the remainder of chr6q. However, nothing precludes clones B and
C from deriving directly from clone A and independently deleting parts of chr6q. It is
interesting that it is clone D, the last clone of the tree, that becomes the dominant clone
in the tumor population according to mixture proportion, suggesting that the events
unique to this clone (e.g. amplification of chr9) may have provided a growth advantage
to this clone.
The clonal allelic profiles for GBM-06-0152 are shown in Figure 5.10. This
tumor is interesting because the amplification of chr7, a characteristic of approximately
40% of GBM tumors [31, 56], does not occur until clone B. Recall that this sample
was shown in the previous chapter to have two double minute chromosomes, one with
94
Chrom
osom
e
Rel
ativ
e C
over
age
Alle
le F
ract
ion
Maj
ority
Alle
le S
tate
Min
ority
Alle
le S
tate
Figure 5.8: Monoclonal allelic profile for GBM-06-0145. The “Relative Coverage” and“Allele Fraction” displayed at the top of the plot shows both the observed output byBamBam and the imputed coverage and allele fraction generated by rendering the dataas the mixture of the single clone and normal contamination. The comparison of realversus modeled data shows very strong agreement. The clone’s allelic profile below showsthe majority and minority allelic states for the tumor genome, showing amplification ofone copy of entire chr7 and chr19, complete loss of one copy of chr10, and arm-level lossof chr9p.
95
Chrom
osom
e
Rel
ativ
e C
over
age
Alle
le F
ract
ion
Maj
ority
Alle
le S
tate
Min
ority
Alle
le S
tate
Maj
ority
Alle
le S
tate
Min
ority
Alle
le S
tate
Maj
ority
Alle
le S
tate
Min
ority
Alle
le S
tate
Maj
ority
Alle
le S
tate
Min
ority
Alle
le S
tate
Figure 5.9: Polyclonal allelic profiles for GBM-06-0185. A total of 4 distinct cloneswere identified in this tumor, with clone D determined to be the dominant clone ofthe population, comprising 42.7% of the tumor sample. All clones have single-copyamplifications of chr7, chr19 & chr20, single copy loss of chr10 and chr22, and loss ofchr9p in common. Clones B, C, & D all have deletions in chr6, but clone B’s deletions arefocal while clones C & D display arm-level loss of chr6q. Clone D is further distinguishedby amplification of the intact copy of chr9.
96
MDM2 & CDK4 and another containing EGFR, that were borne out of a chromothripsis-
like event. While extremely amplified genomic regions are difficult to render in these
clonal allelic profiles, you can see evidence of the deletions related to these events on
chr12 in clone A, suggesting that these double minutes occurred early in this tumor’s
development. It’s possible that the early focal amplification of EGFR may have played
a role in the later emergence of chr7 amplification.
The clonal evolution described by the allelic states for sample GBM-06-1086
has a few interesting aspects. The first subtle thing to notice in its allelic profile, shown
in Figure 5.11, is that the focal deletion of CDKN2A occurs does not occur until clone
B, suggesting it occurred after the complete loss of chr9 observed first in clone A. This is
strong evidence supporting the hypothesis presented in Section 4.3.1 that focal losses of
CDKN2A likely occur after arm-level or entire chromosomal losses of chr9. The second
interesting aspect is that clones C & D exhibit hemizygous losses of 13 different entire
chromosomes. Clone D takes this one step further by deleting its last copy of chr18,
as well as amplifying chr19. This reduces the ploidy of both clones C & D to 1.31,
from the approximately normal ploidy shared by the other two clones (ploidy = 1.95).
It’s fascinating how cells that have lost almost 30% of their genomic content, becoming
haploid for most chromosomes and losing both copies of one chromosome, can not only
survive but, given the 41.8% mixture proportion of clone D, apparently thrive in a
population of tumor cells.
97
Chrom
osom
e
Rel
ativ
e C
over
age
Alle
le F
ract
ion
Maj
ority
Alle
le S
tate
Min
ority
Alle
le S
tate
Maj
ority
Alle
le S
tate
Min
ority
Alle
le S
tate
Maj
ority
Alle
le S
tate
Min
ority
Alle
le S
tate
Figure 5.10: Polyclonal allelic profiles for GBM-06-0152. Three clones were identifiedin this tumor sample that has an estimated 24.1% normal contamination. All clonesshare amplifications of chr1, chr19 & chr20, deletions of chr10 & chr22, and focal lossesof chr12 related to the chromothripsis-like event that created two DMs described inthe previous chapter. Clones B & C exhibit amplification of chr7 and deletions of thenon-amplified copy of chr1 as well as chr2, chr3, chr4, chr8, chr13, and chr17. Clone Cfurther amplifies the remaining copy of chr8.
98
Chr
omos
ome
Rel
ativ
e C
over
age
Alle
le F
ract
ion
Maj
ority
Alle
le S
tate
Min
ority
Alle
le S
tate
Maj
ority
Alle
le S
tate
Min
ority
Alle
le S
tate
Maj
ority
Alle
le S
tate
Min
ority
Alle
le S
tate
Maj
ority
Alle
le S
tate
Min
ority
Alle
le S
tate
Figure 5.11: Polyclonal allelic profiles for GBM-06-1086. Four clones were identified inthis tumor sample that has an estimated 7.5% normal contamination. All clones shareamplification of chr21 and deletions of chr9 & chr11p. Clones C & D exhibit significantchromosomal loss, deleting chr1, chr3, chr4, chr5, chr6, chr8, chr10, chr13, chr14, chr15,chr17, chr18, and chr20. The dominant clone D, making up 41.6% of the tumor sample,further deletes the sole remaining copy of chr18 and amplifies chr19. The black arrowsindicate the position of CDK2NA in clones A & B, highlighting the arrival of the focaldeletion of CDKN2A in the latter clone.
99
5.5.2 Clonality of squamous cell carcinomas of the lung
Whole genome data for nine squamous cell carcinomas of the lung (LUSC)
sequenced by TCGA were analyzed by these methods to infer clonality. The allelic
state diagrams of two tumors are shown in Figure 5.12. From the greater number
of transitional allelic states evident in these two samples, it appears that these LUSC
tumors exhibit a much higher degree of clonality compared to the GBM tumors described
above.
The tumor sample LUSC-66-2756 shown in Figure 5.12(b) exhibits numerous
highly amplified states at ASD vertices (states common among all major clones) and
ASD edges (states shared by only a subset of major clones). A wide variety of mixture
proportions is evident from the almost continuous set of different positions of point clus-
ters along, and in between, ASD edges, suggesting that this sample is highly polyclonal.
Another interesting feature of this sample is that none of its genome is found in the
single copy loss allelic state (1,0). This may have occurred via a genome doubling event
where the tumor genome was briefly tetraploid (N=4), then a series of chromosomal
deletions led to either single copy gain, “normal,” or CN-LOH allelic states. Genome
doubling events are believed to often occur in serous ovarian carcinomas to explain how
large portions of their genomes are observed in the CN-LOH allelic state [15].
100
LUSC-34-2596
LUSC-66-2756
a.
b.
α = 0.33625
α = 0.36375
Majority Allele Fraction
Majority Allele Fraction
Rel
ativ
e C
opy
Num
ber
Rel
ativ
e C
opy
Num
ber
Figure 5.12: Allelic state diagrams for two LUSC tumors: (a) LUSC-34-2596 and (b)LUSC-66-2756. The fitted parameters found normal contamination at 36.4% and 19.6%,respectively. These tumors are highly polyclonal, with LUSC-66-2756 exhibiting a largenumber of highly amplified allelic states and zero single copy loss allelic states, whichcould have resulted from a genome doubling event.
101
5.5.3 Phased mutations to allelic states
To visualize the phased mutations onto allelic states, we will use a slightly
modified allelic state diagram, called the dual allelic state diagram (dual ASD). Noting
from Equation 3.9 that minority allele fraction is the complement of majority allele
fraction (AFmin = 1.0−AFmaj), we can construct a dual ASD by placing a mirror image
of the ASD in order to display the location of the minority allelic states. Mutations
phased to germline variants corresponding to the majority allele, minority allele, or
neither, are plotted on the dual ASD. By determining which allelic state (majority or
minority) the mutations are nearest and using their phased status (if any), we can infer
the the timing of the mutations.
An example dual ASD is shown in Figure 5.13 to help explain how to interpret
it. This dual ASD presents a series of mutations that are phased to germline variants
belonging to either the majority or minority alleles in two different allelic states. Each
mutation’s allele fraction is corrected using the equations presented in Section 5.4.3
and placed onto the dual ASD. Based on their phase and mutant allele fraction, the
dual ASD helps you identify on how many copies of the majority (or minority) alleles
the mutation is present. In the case of the amplification presented in Figure 5.13,
mutations present on both alleles versus those on just one of the amplified copies are
readily distinguished, allowing you to determine visually whether a mutation occurred
before or after the amplification. Similarly, for mutations phased to the minority allele,
we should see them having MAFc ≤ 0.5 for all but the “normal” allelic state where
102
their phased assignment is not meaningful.
Not shown on Figure 5.13 are unphased mutations, where as discussed above,
any time an unphased mutation has MAFc >> 0.5 in unbalanced allelic states, the
mutation must be phased to the majority allele. Otherwise, the mutation exists in
single copy or less, and thus we cannot use its allele fraction to determine whether it’s
phased to the majority or minority alleles.
The dual ASDs for tumor GBM-06-0145 are shown in Figure 5.14. 6 regions
on these diagrams are highlighted to help with interpretation of these diagrams on
real data. Regions (a) and (b) show mutations that were directly phased via nearby
germline variants to the majority allele. Only the 2 mutations in (b) are found to have
the MAFc corresponding to an amplified mutation and hence are expected to have been
present prior to the duplication event (both mutations are located in intergenic regions).
Most majority-phased mutations are found in region (a), corresponding to mutations
at single copy number, suggesting that these mutations occurred post-duplication. An
unphased, missense mutation in DOCK8 is found in the single copy loss allelic state,
meaning that the only copy of DOCK8 remaining in this tumor is in the mutated state.
Inactivation of DOCK8 through homozygous deletion has been linked to progression of
lung cancers, so the lack of wildtype DOCK8 in this GBM tumor may have played a
role in its tumorigenesis [86]. Figure 5.14 also demonstrates the high degree of variation
in estimates of MAFc from these average coverage whole genomes. As TCGA increases
tumor sequencing depth from 30x to 60x, these estimates will improve.8
The most striking thing about the dual ASD for tumor LUSC-34-2596, shown
103
Single Copy Gain
MAFc
Majority Allele
Minority Allele
{“Normal”
Majority Allele Statein Single Copy Loss
Minority Allele Statein Single Copy Loss
Tumor Genome
0.0 0.5 1.0
LOHCN-LOH
1 2 3 4 5 6 7
0.0
2.0
1.0
Rel
ativ
e C
opy
Num
ber
Figure 5.13: Example of seven phased mutations on the dual ASD. A representativeregion on the tumor genome is shown with blue and red horizontal lines representingchromosomes in the tumor mapping to this region, consisting of a subregion in thesingle copy gain allelic state, a subregion in the “normal” allelic state, a subregionexhibiting CN-LOH, and a subregion exhibiting LOH. Mutations 1, 2, and 3 are foundin the amplified subregion, where mutations 2 & 3 are majority-phased (red stars) andmutation 1 is minority-phased (blue star). Mutations 4 & 5 are found in the “normal”allelic state, where mutation 4 is majority-phased and mutation 5 is minority-phased.Both regions exhibiting LOH have one mutation each phased to the sole remaining allele,and thus mutations 6 & 7 are majority-phased. The dual ASD below shows where each ofthese mutations would be found, using each mutation’s corrected allele fraction, MAFc,to determine its location along the x-axis. Note the different placement of mutations2 & 3 in the single-copy gain allelic state, where mutation 3, the mutant allele thatexists on both chromosomes with the majority allele, is found near the single copy gainmajority allelic state. Mutation 2 is found near the single copy gain minority allelicstate, correctly identifying that the mutation exists on only one copy of the majorityallele and hence is likely to occur after the duplication event. Finally, note that theminority-phased mutations in blue are all found in the left-half of the dual ASD.
104
a b
c
d
f
e
DOCK8 (V93M)
MAFc
0.0 0.5 1.0
MAFc
0.0 0.5 1.0 MAFc
0.0 0.5 1.0
Figure 5.14: Phased mutations on the dual ASD for tumor GBM-06-0145. 7 regionsare encircled on these plots: (a) majority-phased mutations in an amplified allelic state,but presenting with MAFc suggesting the mutation is only on one of the two copies, (b)majority-phased mutations in an amplified allelic state with MAFc suggesting muta-tions are present on both amplified copies, (c) majority-phased mutations with an allelefraction consistent with the LOH allelic state, (d) minority-phased mutations in themajority-amplified allelic state with MAFc consistent with single copy, (e) unphasedmutations with MAFc consistent with amplified allelic state, and (f) unphased muta-tions with MAFc consistent with LOH allelic state. LOH has left this cell with only amutated copy of the gene DOCK8.
105
in Figure 5.15, is the sheer number of mutations, phased or unphased, across all expected
allelic states. Compared to the previous GBM tumor, it is clear that the mutation rate
of LUSC-34-2596 is significantly higher. This is expected since lung tumors exhibit
some of the highest mutation rates among the cancers studied thus far by TCGA [34].
We observe a great number of both majority- and minority-phased mutations in the
balanced-amplified allelic state (2,2) at the expected MAFc ≈ 0.5, labeled (a) and (b)
in Figure 5.15. We also observe a large cluster of mutations to the left of the (2, 2)
balanced-amplified state that correspond to mutations in the (3, 1) amplified state at
single copy number. Moving down to the (2, 1) amplified state, as in the GBM sample,
we observe many single copy mutations on the majority allele and hence likely occurring
post-duplication. However, unlike the GBM case, we observe a large cluster of mutations
to the right (present in two copies), representing mutations that were likely there before
the duplication event.
The location of a majority-phased missense mutation in NDRG1, a gene re-
cently discovered to be up-regulated in squamous cell lung cancer [93], is found in a
genomic region in between the “normal” and single copy loss allelic states. Its MAFc
is approximately equal to the allele fraction of the genomic region, suggesting that the
mutation exists on both clones (i.e. the clone with “normal” allelic state and the clone
with single copy loss allelic state). This is evidence that the mutation occurred early,
prior to the emergence of the second clone featuring the new deletion, and that the
deletion contained the wildtype version of NDRG1.
The location of three unphased mutations, BRAF, DNMT3A, and TP53, are
106
also highlighted in Figure 5.15. The nonsense mutation in TP53 is found in a CN-
LOH state and has a mutant allele fraction that precisely corresponds to the CN-LOH
allele fraction, meaning that this tumor has deleted one copy of TP53, knocked out
the remaining copy via mutation, and then amplified the mutant allele. The region
encompassing BRAF was highly amplified, and it is clear from BRAF’s MAFc that
the mutation occurred prior to or early in its amplification. BRAF mutations occur
frequently in melanomas but have been recently discovered in a small percentage of
non-small cell lung carcinomas [20, 12]. Since more than half of the copies are mutated,
the mutation could not have occurred after the amplification process had finished unless
BRAF was independently and identically mutated on multiple copies, a highly improb-
able event. A mutation in DNMT3A, a gene whose loss is implicated in lung cancer
and other tumor types [28, 97, 45], is found in the “normal” allelic state and has the
expected MAFc ≈ 0.5. In all of these cases, mutations to these genes must have oc-
curred early during tumorigenesis as they are present in all (or, in the case of BRAF,
at least the majority) of the tumor’s major clones. Since these are genes known to be
drivers in multiple tumor types, this suggests that one or more of these mutations may
be drivers of this particular tumor.
Table 5.2 summarizes the phaseable mutations for 12 GBM and 8 LUSC tu-
mors. Again, note the higher overall rate of mutations in the LUSC tumors relative to
the GBM tumors. Also, it’s clear that significantly more mutations are found at single
copy within the amplified regions of the GBM tumors, whereas in LUSC tumors we see
no consistent bias, with some tumors having more mutations in single copy and others
107
TP53 (K305*)
BRAF (D380H)
DNMT3A (S878P)
ba
MAFc
0.0 0.5 1.0
MAFc
0.0 0.5 1.0 MAFc
0.0 0.5 1.0
NDRG1(V309M)
Figure 5.15: Phased mutations on the dual ASD for tumor LUSC-34-2596. The twoencircled regions, (a) and (b), show a number of mutations phased to the majority andminority alleles, respectively, in the balanced amplified allelic state (2,2). One majority-phased mutation in NDRG1 is found in a transitional allelic state with matching MAFc.The locations of two missense mutations, in BRAF & DNMT3A, and one nonsensemutation in TP53 are shown in the unphased plot, placing BRAF in a highly amplifiedallelic state, DNMT3A in the “normal” allelic state, and TP53 in the CN-LOH state.
108
having more mutations in double copy (or more). Assuming the mutation rate remains
constant throughout the development of these tumors, then this suggest that amplifi-
cation usually occurs early in the development of the GBM tumors, before most other
mutations occur, whereas in LUSC tumor development, amplification may occur at any
stage. Mutations and copy number alterations are frequent occurrences of LUSC tumor
development, with large numbers of mutations occurring prior to and after amplification
events.
Another possibility is that the mutation rate does not remain constant during
development of GBM tumors. Suppose that amplification of the growth factor EGFR,
a common event in these GBM tumors, increases the cell’s rate of growth and subse-
quently reduces the cell’s ability to correct mistakes made during genome replication,
thereby increasing the mutation rate per cellular division. This would also explain the
enrichment of mutations present in single copy within amplified allelic states. However,
without knowledge of the number of generations that occur before and after EGFR
amplification, we cannot determine if the effect we observe is due to timing bias or
mutation rate change.
5.6 Discussion
The methods presented here demonstrate our ability to computationally dis-
sect a tumor’s cell population using whole genome sequencing data. We present an
effective way to visually assess a tumor sample’s clonality using an allelic state diagram.
109
Table 5.2: Phased mutations in GBM and LUSC tumors.
Sample Total Maj-Phased In Amp. Single Copy Min-Phased
GBM-06-0188 1,595 67 2 1 (50%) 19GBM-06-0648 1,600 68 9 5 (56%) 34GBM-06-0877 4,792 367 130 84 (65%) 145GBM-14-1454 2,539 291 55 38 (69%) 80GBM-06-0152 2,896 269 47 35 (74%) 80GBM-26-1438 2,820 239 36 32 (89%) 86GBM-06-0155 7,167 1,108 99 93 (94%) 440GBM-06-0214 2,827 237 31 29 (94%) 83GBM-06-0145 3,800 511 56 53 (95%) 225GBM-14-0786 5,863 783 117 114 (97%) 241GBM-06-1086 3,993 412 10 10 (100%) 129GBM-06-0185 1,195 20 2 2 (100%) 14
Total GBM 41,087 4,372 594 496 (84%) 1,576
LUSC-66-2756 45,995 7,336 6,869 2,107 (31%) 2,104LUSC-56-1622 19,343 3,159 286 99 (35%) 1,016LUSC-60-2695 6,265 830 408 148 (36%) 302LUSC-43-3394 18,905 4,162 694 267 (38%) 1,184LUSC-66-2757 11,400 1,477 374 177 (47%) 498LUSC-60-2713 21,337 4,290 503 262 (52%) 1,207LUSC-34-2596 30,794 7,166 663 349 (53%) 1,410LUSC-60-2722 32,485 5,101 504 319 (63%) 1,682
Total LUSC 186,524 33,521 10,301 3,728 (36%) 9,403Total LUSC* 140,529 26,185 3,432 1,621 (47%) 7,299
Total = Number of phased and unphased mutations called, Maj-Phased = Number
of majority-phased mutations, In Amp. = Number of majority-phased mutations in
regions of amplified allelic state, Single Copy = Number of majority-phased mutations
in regions of amplified allelic state, but havingMAFc corresponding to single copy, Min-
Phased = Number of minority-phased mutations. Total LUSC* is the total for all
LUSC tumors excluding the outlier case of LUSC-66-2756 that has almost double the
number of mutations of any other LUSC tumor.
110
We show how the clonal mixture of a tumor can be determined, decomposing the tumor
population into all of the major clones of the tumor cell population and estimate normal
contamination in order to fully account for the copy number and allele fraction data
measured by BamBam. We further show that whole genome allelic profiles of all major
clones can be fully determined and phased, allowing us to infer the phylogenetic tree of
polyclonal tumor genomes to time the emergence of clone-specific copy number alter-
ations. Finally, by using phasing and mutant allele fraction, we can time the emergence
of mutations with respect to copy number alterations of the chromosomal region that
carry them.
The clonality and timing information helps us gain a better appreciation for
the dynamic nature of individual tumors. We observe tumors belonging to the same
tumor type exhibiting a wide range of clonality, from monoclonal to highly polyclonal.
We also find that certain tumor types reflect an overall higher degree of clonality than
others, possibly suggesting that capacity to nurture multiple clones is in some way
tissue-dependent. Most importantly, all of this information can be discovered from
just a single tumor biopsy, making such methods feasible for future applications in the
clinical environment, as will be further discussed below in Section 6.2.
111
Chapter 6
Conclusion
In this thesis, numerous whole genome sequencing analysis methods were de-
scribed and used on sequencing data from real tumors to prove their individual and col-
lective efficacies. Starting with BamBam and Bridget, we discover the full complement
of germline and somatic variants from the very small alterations (e.g. point mutations)
to large structural events (e.g. deletions of entire chromosomes), and everything in be-
tween. While this knowledge is powerful in isolation, through clever integration of these
data we gain a much truer understanding of the tumor.
In this concluding chapter, I will briefly describe some of the directions I foresee
this work taking in both research and clinical arenas.
112
6.1 Clone-directed mutation discovery
By determining the clonal mixture, we could determine the expected mutant
allele fraction for mutations specific to each clone in the tumor. This information could
be used to adjust the mutation model used by BamBam, enabling confident calls to be
made at lower mutation allele frequencies than they would otherwise given the prior
knowledge that clones exist in the sample that should produce mutations with certain
mutant allele frequencies.
There is a lower limit to our ability to detect variants that is dependent on the
coverage of the tumor genome. For example, from a 30x tumor whole genome with 20%
normal contamination, we expect only 5 reads on average to come from a clone that
comprises 20% of the tumor population (assuming a “normal” allelic state). Since most
mutations are found on just one of the two parental alleles, only 2-3 of those reads are
likely to exhibit the mutation. Therefore, increased coverage, possibly by integrating
high coverage exomes (200-1000x), would be necessary to yield sufficient coverage to
identify these clone-specific mutations.
The tweak to the current algorithm suggested above would be a good first
step to extracting clone-specific mutations, but a mutation model that incorporates the
phylogenetic relationships between clones may be much more powerful. This method
would help find mutations that are specific to each clone, but would also be able to infer
the node in the tree where the mutation likely emerged, to better model the situation
where multiple clones may share the mutation. A phylogenetic-based mutation model
113
could also prove useful to analyze the mutations of related samples, such as a primary
tumor and its metastases, to reconstruct the mutational history of a cancer as it spreads.
6.2 Clinical applications
Presently, there is much hype surrounding the promise of genome sequencing
in improving the care and treatment of cancer. The hope is that with improved knowl-
edge about the unique features of each tumor genome, physicians will be able to tailor
treatment to more precisely target the cancer. Most efforts focus on specific oncogenic
mutations, or genes commonly amplified in tumors, and therefore employ targeted se-
quencing (e.g. exome capture, or even more restrictive kits that capture ∼200 cancer
genes targeted by drugs that either FDA-approved or in late-phase clinical trials) [24].
Less effort is applied to whole genome sequencing data since it currently generates a
huge amount of data, most of which is likely to go unused as its relevance to cancer
treatment is unknown or non-existent. However, the methods described in this thesis
clearly suggest some novel clinical applications of whole genome sequencing data.
By integrating copy number and structural variants in regions exhibiting ex-
treme amplification, double minute chromosomes (DM) were identified and reconstructed
to determine their genetic content, accurately down to the base level. DNA reads from
these DMs, born in a brain tumor, were readily detectable in the bloodstream of the
same patients. This has obvious clinical relevance, as it suggests that brain tumors could
be detected or monitored by a simple blood draw. If the creation of a DM is an early or
114
initiating event of tumorigenesis, then it’s plausible that a sensitive blood-based assay
could be developed to detect certain brain tumors even before they could be detected
by more expensive means, such as an MRI.
One can envision a new class of sequencing-based assays that monitor DNA
from blood samples to determine, in the de novo sense, if a region surrounding any
known oncogene, such as EGFR, c-Myc, and MDM2, exhibit a clustered pattern of
breakpoints indicative of an amplified DM. Collecting evidence of discordant reads across
regions that have oncogenic potential should improve our ability to identify these regions
even when the concentration of DMs in the bloodstream is low.
The ability to detect DMs in the bloodstream should extend to other can-
cers that commonly feature highly amplified DMs containing known oncogenes, such as
EGFR in non-small cell lung cancer and c-Myc in acute myelogenous leukemia [89, 78].
In fact, our ability to detect DMs with this method may improve for tumor types where
the bloodstream is more accessible to the tumor. Furthermore, drugs that specifically
target genes commonly amplified in DMs may be prescribed based on evidence collected
solely from the bloodstream, avoiding painful and dangerous tumor biopsies. Future
studies on a wider panel of tumors and tumor types will help us learn how well this
method can be extended to other cancers.
In my opinion, tumor clonality has even more potential in the clinic. The
ability to determine a tumor’s clonality and identify all of the major clones that comprise
the growing tumor mass, all from the whole genome sequencing data of a single biopsy,
opens up a wide variety of potential clinical applications. For example, suppose a
115
newly-diagnosed patient’s tumor is biopsied and all of the major clones are identified
via clonal analysis. The clinician could use this clonality analysis to tailor the patient’s
treatment according to the alterations specific to the clone furthest up the evolutionary
tree, the progenitor tumor cell, with the hope that by treating the initial tumor mass,
all derivative clones will also be targeted.
In a related twist, suppose a patient is diagnosed with a slow-growing tumor
that can be safely monitored for a longer period of time before surgery or the beginning
of chemotherapy. Clonal analysis of a series of biopsies could be performed and, by
tracking the clonal composition of the tumor through time, the clinician could identify
the clones that are growing most rapidly. By designing a treatment that targets not
what is currently the dominant clone, but the clone that is becoming the dominant
clone, we might more effectively treat the cancer.
Clonality analysis may also prove useful in better understanding the metastatic
spread of cancers. A clonal analysis of a primary tumor and a series of metastases
could determine all of the major clones present in the spreading tumor. By inspecting
the clonal composition of the primary tumor and each metastasis, we could determine
how each clone spreads and discover if there are particular clones that exhibit superior
metastatic potential. If we can determine the characteristics unique to the metastatic
clones, it may be possible to identify the emergence of these characteristics in the
minor clones of another patient’s primary tumor, providing an “early warning” for
the likelihood of imminent metastasis.
116
6.3 Final thoughts
While this thesis presents a comprehensive range of techniques to explore and
understand the cancer genome, I know the work presented here merely represents the
“tip of the iceberg” of what is to be learned about cancer genome. The paradoxical
nature of cancer research is such that, for every bit of information discovered, we both
add knowledge and yet reduce our overall understanding by revealing further complex-
ities of this disease. One day soon the balance of knowledge will tip toward our favor,
and the days when cancer is considered a death sentence will be behind us.
117
Bibliography
[1] ACS. Cancer Facts & Figures 2012. American Cancer Society, 2012.
[2] Stephen F Altschul, Warren Gish, Webb Miller, Eugene W Myers, and David J Lip-
man. Basic local alignment search tool. Journal of Molecular Biology, 215(3):403–
410, October 1990.
[3] Stylianos E Antonarakis, Michael Krawczak, and David N Cooper. Disease-causing
mutations in the human genome. European Journal of Pediatrics, 159(15):S173–
S178, January 2000.
[4] M Ashburner, C A Ball, J A Blake, D Botstein, H Butler, J M Cherry, A P
Davis, K Dolinski, S S Dwight, J T Eppig, M A Harris, D P Hill, L Issel-Tarver,
A Kasarskis, S Lewis, J C Matese, J E Richardson, M Ringwald, G M Rubin, and
G Sherlock. Gene ontology: tool for the unification of biology. The Gene Ontology
Consortium. Nature genetics, 25(1):25–29, May 2000.
[5] Leonora Balaj, Ryan Lessard, Lixin Dai, Yoon-Jae Cho, Scott L Pomeroy, Xan-
dra O Breakefield, and Johan Skog. Tumour microvesicles contain retrotransposon
118
elements and amplified oncogene sequences. Nature Communications, 2:180, Febru-
ary 2011.
[6] P E Barker. Double minutes in human tumor cells. Cancer Genetics and Cytoge-
netics, 5(1):81–94, February 1982.
[7] F Baskin, R N Rosenberg, and V Dev. Correlation of double-minute chromosomes
with unstable multidrug cross-resistance in uptake mutants of neuroblastoma cells.
Proceedings of the National Academy of Sciences, 78(6):3654–3658, January 1981.
[8] Y Ben-Neriah, G Q Daley, A M Mes-Masson, O N Witte, and D Baltimore. The
chronic myelogenous leukemia-specific P210 protein is the product of the bcr/abl
hybrid gene. Science, 233(4760):212–214, July 1986.
[9] J Bentley. DNA double strand break repair in human bladder cancer is error
prone and involves microhomology-associated end-joining. Nucleic Acids Research,
32(17):5249–5259, September 2004.
[10] Rameen Beroukhim, Gad Getz, Leia Nghiemphu, Jordi Barretina, Teli Hsueh,
David Linhart, Igor Vivanco, Jeffrey C Lee, Julie H Huang, Sethu Alexander,
Jinyan Du, Tweeny Kau, Roman K Thomas, Kinjal Shah, Horacio Soto, Sven
Perner, John Prensner, Ralph M Debiasi, Francesca Demichelis, Charlie Hatton,
Mark A Rubin, Levi A Garraway, Stan F Nelson, Linda Liau, Paul S Mischel,
Tim F Cloughesy, Matthew Meyerson, Todd A Golub, Eric S Lander, Ingo K
Mellinghoff, and William R Sellers. Assessing the significance of chromosomal
119
aberrations in cancer: methodology and application to glioma. Proc Natl Acad Sci
USA, 104(50):20007–20012, December 2007.
[11] I Bieche, C Tomasetto, C H Regnier, C Moog-Lutz, M C Rio, and R Lidereau. Two
distinct amplified regions at 17q11-q21 involved in human primary breast cancer.
Cancer Res, 56(17):3886–3890, September 1996.
[12] Marcia S Brose, Patricia Volpe, Michael Feldman, Madhu Kumar, Irum Rishi,
Renee Gerrero, Eugene Einhorn, Meenhard Herlyn, John Minna, Andrew Nichol-
son, Jack A Roth, Steven M Albelda, Helen Davies, Charles Cox, Graham Brignell,
Philip Stephens, P Andrew Futreal, Richard Wooster, Michael R Stratton, and Bar-
bara L Weber. BRAF and RAS mutations in human lung cancer and melanoma.
Cancer Research, 62(23):6997–7000, December 2002.
[13] Peter J Campbell, Erin D Pleasance, Philip J Stephens, Ed Dicks, Richard
Rance, Ian Goodhead, George A Follows, Anthony R Green, P Andy Futreal,
and Michael R Stratton. Subclonal phylogenetic structures in cancer revealed by
ultra-deep sequencing. Proc Natl Acad Sci USA, 105(35):13081–13086, September
2008.
[14] S M Carroll, M L DeRose, P Gaudray, C M Moore, D R Needham-Vandevanter,
D D Von Hoff, and G M Wahl. Double minute chromosomes can be produced from
precursors derived from a chromosomal deletion. Molecular and Cellular Biology,
8(4):1525–1533, April 1988.
120
[15] Scott L Carter, Kristian Cibulskis, Elena Helman, Aaron McKenna, Hui Shen,
Travis Zack, Peter W Laird, Robert C Onofrio, Wendy Winckler, Barbara A Weir,
Rameen Beroukhim, David Pellman, Douglas A Levine, Eric S Lander, Matthew
Meyerson, and Gad Getz. Absolute quantification of somatic DNA alterations in
human cancer. Nature, 30(5):413–421, April 2012.
[16] James A Goodrich Robert Tjian Catherine J Thut. Repression of p53-mediated
transcription by MDM2: adualmechanism. Genes & Development, 11(15):1974,
August 1997.
[17] Michael A Chapman, Michael S Lawrence, Jonathan J Keats, Kristian Cibulskis,
Carrie Sougnez, Anna C Schinzel, Christina L Harview, Jean-Philippe Brunet, Gre-
gory J Ahmann, Mazhar Adli, Kenneth C Anderson, Kristin G Ardlie, Daniel Au-
clair, Angela Baker, P Leif Bergsagel, Bradley E Bernstein, Yotam Drier, Rafael
Fonseca, Stacey B Gabriel, Craig C Hofmeister, Sundar Jagannath, Andrzej J
Jakubowiak, Amrita Krishnan, Joan Levy, Ted Liefeld, Sagar Lonial, Scott Mahan,
Bunmi Mfuko, Stefano Monti, Louise M Perkins, Robb Onofrio, Trevor J Pugh,
S Vincent Rajkumar, Alex H Ramos, David S Siegel, Andrey Sivachenko, A Keith
Stewart, Suzanne Trudel, Ravi Vij, Douglas Voet, Wendy Winckler, Todd Zim-
merman, John Carpten, Jeff Trent, William C Hahn, Levi A Garraway, Matthew
Meyerson, Eric S Lander, Gad Getz, and Todd R Golub. Initial genome sequencing
and analysis of multiple myeloma. Nature, 471(7339):467–472, March 2011.
[18] S H Chen and C R Scott. Recombination between two 14-bp homologous sequences
121
as the mechanism for gene deletion in factor IX Seattle 1. Am J Hum Genet,
47(6):1020–1022, December 1990.
[19] E D Coene, V Schelfhout, R A Winkler, A M Schelfhout, N Van Roy, M Groote-
claes, F Speleman, and C R De Potter. Amplification units and translocation at
chromosome 17q and c-erbB-2 overexpression in the pathogenesis of breast cancer.
Virchows Arch, 430(5):365–372, May 1997.
[20] Helen Davies, Graham R Bignell, Charles Cox, Philip Stephens, Sarah Edkins,
Sheila Clegg, Jon Teague, Hayley Woffendin, Mathew J Garnett, William Bottom-
ley, Neil Davis, Ed Dicks, Rebecca Ewing, Yvonne Floyd, Kristian Gray, Sarah
Hall, Rachel Hawes, Jaime Hughes, Vivian Kosmidou, Andrew Menzies, Cather-
ine Mould, Adrian Parker, Claire Stevens, Stephen Watt, Steven Hooper, Rebecca
Wilson, Hiran Jayatilake, Barry A Gusterson, Colin Cooper, Janet Shipley, Dar-
ren Hargrave, Katherine Pritchard-Jones, Norman Maitland, Georgia Chenevix-
Trench, Gregory J Riggins, Darell D Bigner, Giuseppe Palmieri, Antonio Cossu,
Adrienne Flanagan, Andrew Nicholson, Judy W C Ho, Suet Y Leung, Siu T
Yuen, Barbara L Weber, Hilliard F Seigler, Timothy L Darrow, Hugh Paterson,
Richard Marais, Christopher J Marshall, Richard Wooster, Michael R Stratton,
and P Andrew Futreal. Mutations of the BRAF gene in human cancer. Nature,
417(6892):949–954, June 2002.
[21] N D Dees, Q Zhang, C Kandoth, M C Wendl, W Schierding, D C Koboldt, T B
Mooney, M B Callaway, D Dooling, E R Mardis, R K Wilson, and L Ding. Mu-
122
SiC: Identifying mutational significance in cancer genomes. Genome Research,
22(8):1589–1598, August 2012.
[22] M F Denissenko, A Pao, G P Pfeifer, and M Tang. Slow repair of bulky DNA
adducts along the nontranscribed strand of the human p53 gene may explain the
strand bias of transversion mutations in cancers. Oncogene, 16(10):1241–1247,
March 1998.
[23] B J Druker, M Talpaz, D J Resta, B Peng, E Buchdunger, J M Ford, N B Lydon,
H Kantarjian, R Capdeville, S Ohno-Jones, and C L Sawyers. Efficacy and safety
of a specific inhibitor of the BCR-ABL tyrosine kinase in chronic myeloid leukemia.
N Engl J Med, 344(14):1031–1037, April 2001.
[24] Michael Eisenstein. Foundation Medicine. Nature, 30(1):14–14, January 2012.
[25] B Elliott and M Jasin. Double-strand breaks and translocations in cancer. Cellular
and molecular life sciences : CMLS, 59(2):373–385, February 2002.
[26] Bernard M Fine, Martin Stanulla, Martin Schrappe, Minh Ho, Susanne Viehmann,
Jochen Harbott, and Linda M Boxer. Gene expression patterns associated with
recurrent chromosomal translocations in acute lymphoblastic leukemia. Blood,
103(3):1043–1049, February 2004.
[27] Stefan Frohling, Claudia Scholl, Ross L Levine, Marc Loriaux, Titus J Boggon,
Olivier A Bernard, Roland Berger, Hartmut Dohner, Konstanze Dohner, Ben-
jamin L Ebert, Sewit Teckie, Todd R Golub, Jingrui Jiang, Marcus M Schitten-
123
helm, Benjamin H Lee, James D Griffin, Richard M Stone, Michael C Heinrich,
Michael W Deininger, Brian J Druker, and D Gary Gilliland. Identification of driver
and passenger mutations of FLT3 by high-throughput DNA sequence analysis and
functional assessment of candidate alleles. Cancer Cell, 12(6):501–513, December
2007.
[28] Qing Gao, Eveline J Steine, M Inmaculada Barrasa, Dirk Hockemeyer, Mathias
Pawlak, Dongdong Fu, Seshamma Reddy, George W Bell, and Rudolf Jaenisch.
Deletion of the de novo DNA methyltransferase Dnmt3a promotes lung tumor
progression. Proceedings of the National Academy of Sciences of the United States
of America, 108(44):18061–18066, November 2011.
[29] R Goya, M G F Sun, R D Morin, G Leung, G Ha, K C Wiegand, J Senz, A Crisan,
M A Marra, M Hirst, D Huntsman, K P Murphy, S Aparicio, and S P Shah.
SNVMix: predicting single nucleotide variants from next-generation sequencing of
tumors. Bioinformatics, 26(6):730–736, March 2010.
[30] C D Greenman, E D Pleasance, S Newman, F Yang, B Fu, S Nik-Zainal, D Jones,
K W Lau, N Carter, P A W Edwards, P A Futreal, M R Stratton, and P J
Campbell. Estimation of rearrangement phylogeny for cancer genomes. Genome
Research, 22(2):346–361, February 2012.
[31] W J Gullick. Prevalence of aberrant expression of the epidermal growth factor
receptor in human cancers. British Medical Bulletin, pages 87–98, January 1991.
124
[32] P Hainaut. Patterns of p53 G-¿T transversions in lung cancers reflect the primary
mutagenic signature of DNA-damage by tobacco smoke. Carcinogenesis, 22(3):367–
374, March 2001.
[33] P Hainaut, T Hernandez, A Robinson, P Rodriguez-Tome, T Flores, M Hollstein,
C C Harris, and R Montesano. IARC Database of p53 gene mutations in human
tumors and cell lines: updated compilation, revised formats and new visualisation
tools. Nucleic Acids Res, 26(1):205–213, January 1998.
[34] Peter S Hammerman, D Neil Hayes, Matthew D Wilkerson, Nikolaus Schultz, Ron
Bose, Andy Chu, Eric A Collisson, Leslie Cope, Chad J Creighton, Gad Getz,
James G Herman, Bruce E Johnson, Raju Kucherlapati, Marc Ladanyi, Christo-
pher A Maher, Gordon Robertson, Chris Sander, Ronglai Shen, Rileen Sinha,
Andrey Sivachenko, Roman K Thomas, William D Travis, Ming-Sound Tsao,
John N Weinstein, Dennis A Wigle, Stephen B Baylin, Ramaswamy Govindan,
and Matthew Meyerson. Comprehensive genomic characterization of squamous
cell lung cancers. Nature, 489(7417):519–525, September 2012.
[35] D Hanahan and R A Weinberg. The hallmarks of cancer. Cell, 100(1):57–70,
January 2000.
[36] Douglas Hanahan and Robert A Weinberg. Hallmarks of cancer: the next genera-
tion. Cell, 144(5):646–674, March 2011.
[37] A John Iafrate, Lars Feuk, Miguel N Rivera, Marc L Listewnik, Patricia K Donahoe,
125
Ying Qi, Stephen W Scherer, and Charles Lee. Detection of large-scale variation
in the human genome. Nature genetics, 36(9):949–951, September 2004.
[38] Minoru Kanehisa, Michihiro Araki, Susumu Goto, Masahiro Hattori, Mika Hi-
rakawa, Masumi Itoh, Toshiaki Katayama, Shuichi Kawashima, Shujiro Okuda,
Toshiaki Tokimatsu, and Yoshihiro Yamanishi. KEGG for linking genomes to life
and the environment. Nucleic Acids Res, 36(Database issue):D480–4, January 2008.
[39] K K Khanna and S P Jackson. DNA double-strand breaks: signaling, repair and
the cancer connection. Nature genetics, 27(3):247–254, March 2001.
[40] Wigard P Kloosterman, Marlous Hoogstraat, Oscar Paling, Masoumeh Tavakoli-
Yaraki, Ivo Renkens, Joost S Vermaat, Markus J van Roosmalen, Stef van Lieshout,
Isaac J Nijman, Wijnand Roessingh, Ruben van ’t Slot, Jose van de Belt, Victor
Guryev, Marco Koudijs, Emile Voest, and Edwin Cuppen. Chromothripsis is a
common mechanism driving genomic rearrangements in primary and metastatic
colorectal cancer. Genome Biol, 12(10):R103, 2011.
[41] Chandan Kumar-Sinha, Scott A Tomlins, and Arul M Chinnaiyan. Evidence of
recurrent gene fusions in common epithelial tumors. Trends in molecular medicine,
12(11):529–536, November 2006.
[42] D E Larson, C C Harris, K Chen, D C Koboldt, T E Abbott, D J Dooling, T J Ley,
E R Mardis, R K Wilson, and L Ding. SomaticSniper: identification of somatic
126
point mutations in whole genome sequencing data. Bioinformatics, 28(3):311–317,
January 2012.
[43] C Latham, A Zhang, A Nalbanti, S Maner, P Zickert, H Blegen, and A Zetter-
berg. Frequent co-amplification of two different regions on 17q in aneuploid breast
carcinomas. Cancer Genet Cytogenet, 127(1):16–23, May 2001.
[44] I Lavon, D Zrihan, B Zelikovitch, Y Fellig, D Fuchs, D Soffer, and T Siegal. Lon-
gitudinal Assessment of Genetic and Epigenetic Markers in Oligodendrogliomas.
Clinical Cancer Research, 13(5):1429–1437, March 2007.
[45] Timothy J Ley, Li Ding, Matthew J Walter, Michael D McLellan, Tamara Lam-
precht, David E Larson, Cyriac Kandoth, Jacqueline E Payton, Jack Baty, John
Welch, Christopher C Harris, Cheryl F Lichti, R Reid Townsend, Robert S Fulton,
David J Dooling, Daniel C Koboldt, Heather Schmidt, Qunyuan Zhang, John R
Osborne, Ling Lin, Michelle O’Laughlin, Joshua F McMichael, Kim D Delehaunty,
Sean D McGrath, Lucinda A Fulton, Vincent J Magrini, Tammi L Vickery, Jasreet
Hundal, Lisa L Cook, Joshua J Conyers, Gary W Swift, Jerry P Reed, Patricia A
Alldredge, Todd Wylie, Jason Walker, Joelle Kalicki, Mark A Watson, Sharon
Heath, William D Shannon, Nobish Varghese, Rakesh Nagarajan, Peter Wester-
velt, Michael H Tomasson, Daniel C Link, Timothy A Graubert, John F DiPersio,
Elaine R Mardis, and Richard K Wilson. DNMT3AMutations in Acute Myeloid
Leukemia. N Engl J Med, 363(25):2424–2433, December 2010.
[46] Heng Li, Bob Handsaker, Alec Wysoker, Tim Fennell, Jue Ruan, Nils Homer, Gabor
127
Marth, Goncalo Abecasis, Richard Durbin, and 1000 Genome Project Data Pro-
cessing Subgroup. The Sequence Alignment/Map format and SAMtools. Bioinfor-
matics, 25(16):2078–2079, August 2009.
[47] Heng Li, Jue Ruan, and Richard Durbin. Mapping short DNA sequencing reads and
calling variants using mapping quality scores. Genome Research, 18(11):1851–1858,
November 2008.
[48] Marc M Loriaux, Ross L Levine, Jeffrey W Tyner, Stefan Frohling, Claudia Scholl,
Eric P Stoffregen, Gerlinde Wernig, Heidi Erickson, Christopher A Eide, Roland
Berger, Olivier A Bernard, James D Griffin, Richard M Stone, Benjamin Lee,
Matthew Meyerson, Michael C Heinrich, Michael W Deininger, D Gary Gilliland,
and Brian J Druker. High-throughput sequence analysis of the tyrosine kinome in
acute myeloid leukemia. Blood, 111(9):4788–4796, May 2008.
[49] Elaine R Mardis, Li Ding, David J Dooling, David E Larson, Michael D McLellan,
Ken Chen, Daniel C Koboldt, Robert S Fulton, Kim D Delehaunty, Sean D Mc-
Grath, Lucinda A Fulton, Devin P Locke, Vincent J Magrini, Rachel M Abbott,
Tammi L Vickery, Jerry S Reed, Jody S Robinson, Todd Wylie, Scott M Smith,
Lynn Carmichael, James M Eldred, Christopher C Harris, Jason Walker, Joshua B
Peck, Feiyu Du, Adam F Dukes, Gabriel E Sanderson, Anthony M Brummett, Eric
Clark, Joshua F McMichael, Rick J Meyer, Jonathan K Schindler, Craig S Pohl,
John W Wallis, Xiaoqi Shi, Ling Lin, Heather Schmidt, Yuzhu Tang, Carrie Haipek,
Madeline E Wiechert, Jolynda V Ivy, Joelle Kalicki, Glendoria Elliott, Rhonda E
128
Ries, Jacqueline E Payton, Peter Westervelt, Michael H Tomasson, Mark A Wat-
son, Jack Baty, Sharon Heath, William D Shannon, Rakesh Nagarajan, Daniel C
Link, Matthew J Walter, Timothy A Graubert, John F DiPersio, Richard K Wilson,
and Timothy J Ley. Recurring mutations found by sequencing an acute myeloid
leukemia genome. N Engl J Med, 361(11):1058–1066, September 2009.
[50] A W McPherson, C Wu, A Wyatt, S P Shah, C Collins, and S C Sahinalp. nFuse:
Discovery of complex genomic rearrangements in cancer using high-throughput
sequencing. Genome Research, June 2012.
[51] Mitch McVey and Sang Eun Lee. MMEJ repair of double-strand breaks (director’s
cut): deleted sequences and alternative endings. Cell, 24(11):529–538, November
2008.
[52] Paula L Miliani de Marval, Everardo Macias, Claudio J Conti, and Marcelo L
Rodriguez-Puebla. Enhanced malignant tumorigenesis in Cdk4 transgenic mice.
Nature, 23(10):1863–1873, December 2003.
[53] Felix Mitelman, Bertil Johansson, and Fredrik Mertens. Fusion genes and rear-
ranged genes as a linear function of chromosome aberrations in cancer. Nature
genetics, 36(4):331–334, April 2004.
[54] J K Moore and J E Haber. Cell cycle and genetic requirements of two pathways
of nonhomologous end-joining repair of double-strand breaks in Saccharomyces
cerevisiae. Molecular and Cellular Biology, 16(5):2164–2173, May 1996.
129
[55] David K Moscatello, Marina Holgado-Madruga, Andrew K Godwin, Gloria
Ramirez, George Gunn, Philip W Zoltick, Jaclyn A Biegel, Roberta L Hayes, and
Albert J Wong. Frequent Expression of a Mutant Epidermal Growth Factor Re-
ceptor in Multiple Human Tumors. cancerres.aacrjournals.org.
[56] Cancer Genome Atlas Research Network. Comprehensive genomic characterization
defines human glioblastoma genes and core pathways. Nature, 455(7216):1061–
1068, October 2008.
[57] Cancer Genome Atlas Research Network. Integrated genomic analyses of ovarian
carcinoma. Nature, 474(7353):609–615, June 2011.
[58] G P Nielsen, A O Stemmer-Rachamimov, Y Ino, M B Moller, A E Rosenberg, and
D N Louis. Malignant transformation of neurofibromas in neurofibromatosis 1 is
associated with CDKN2A/p16 inactivation. The American Journal of Pathology,
155(6):1879–1884, December 1999.
[59] Serena Nik-Zainal, Peter Van Loo, David C Wedge, Ludmil B Alexandrov, Christo-
pher D Greenman, King Wai Lau, Keiran Raine, David Jones, John Marshall,
Manasa Ramakrishna, Adam Shlien, Susanna L Cooke, Jonathan Hinton, An-
drew Menzies, Lucy A Stebbings, Catherine Leroy, Mingming Jia, Richard Rance,
Laura J Mudie, Stephen J Gamble, Philip J Stephens, Stuart McLaren, Patrick S
Tarpey, Elli Papaemmanuil, Helen R Davies, Ignacio Varela, David J McBride,
Graham R Bignell, Kenric Leung, Adam P Butler, Jon W Teague, Sancha Martin,
Goran Jonsson, Odette Mariani, Sandrine Boyault, Penelope Miron, Aquila Fa-
130
tima, Anita Langerød, Samuel A J R Aparicio, Andrew Tutt, Anieta M Sieuwerts,
Ake Borg, Gilles Thomas, Anne Vincent Salomon, Andrea L Richardson, Anne-
Lise Børresen-Dale, P Andrew Futreal, Michael R Stratton, and Peter J Campbell.
The Life History of 21 Breast Cancers. Cell, 149(5):994–1007, May 2012.
[60] Tsutomu Nobori, Kaoru Miura, David J Wu, Augusto Lois, Kenji Takabayashi,
and Dennis A Carson. Deletions of the cyclin-dependent kinase-4 inhibitor gene in
multiple human cancers. Nature, 368(6473):753–756, April 1994.
[61] Hitoshi Okada and Tak W Mak. Pathways of apoptotic and non-apoptotic death
in tumour cells. Nat Rev Cancer, 4(8):592–603, August 2004.
[62] C M Perou, T Sørlie, M B Eisen, M van de Rijn, S S Jeffrey, C A Rees, J R Pollack,
D T Ross, H Johnsen, L A Akslen, O Fluge, A Pergamenschikov, C Williams, S X
Zhu, P E Lønning, A L Børresen-Dale, P O Brown, and D Botstein. Molecular
portraits of human breast tumours. Nature, 406(6797):747–752, August 2000.
[63] A Roth, J Ding, R Morin, A Crisan, G Ha, R Giuliany, A Bashashati, M Hirst,
G Turashvili, A Oloumi, M A Marra, S Aparicio, and S P Shah. JointSNVMix: a
probabilistic model for accurate detection of somatic mutations in normal/tumour
paired next-generation sequencing data. Bioinformatics, 28(7):907–913, March
2012.
[64] Astrid M Roy-Engel. LINEs, SINEs and other retroelements: do birds of a feather
flock together? Frontiers in Bioscience, 17:1345.
131
[65] V Ling C P Stanners S M Robertson. Co-amplification of double minute chromo-
somes, multiple drug resistance, and cell surface P-glycoprotein in DNA-mediated
transformants of mouse cells. Molecular and Cellular Biology, 4(3):500, March
1984.
[66] Giuseppe Saglio, Alessandro Morotti, Giovanna Mattioli, Emanuela Messa, Emilia
Giugliano, Gisella Volpe, Giovanna Rege-Cambrin, and Daniela Cilloni. Rational
approaches to the design of therapeutics targeting molecular markers: the case of
chronic myelogenous leukemia. Ann N Y Acad Sci, 1028:423–431, December 2004.
[67] J Z Sanborn, S C Benz, B Craft, C Szeto, K M Kober, L Meyer, C J Vaske, M Gold-
man, K E Smith, R M Kuhn, D Karolchik, W J Kent, J M Stuart, D Haussler, and
J Zhu. The UCSC cancer genomics browser: update 2011. Nucleic Acids Research,
39(Database):D951–D959, December 2010.
[68] Carl F Schaefer, Kira Anthony, Shiva Krupa, Jeffrey Buchoff, Matthew Day, Timo
Hannay, and Kenneth H Buetow. PID: the Pathway Interaction Database. Nucleic
Acids Res, 37(Database issue):D674–9, January 2009.
[69] T Schepeler, P Lamy, J R Laurberg, N Fristrup, T Reinert, J Bartkova, L Tropia,
J Bartek, T D Halazonetis, C-C Pan, M Borre, L Dyrskjøt, and T F Ørntoft. A
high resolution genomic portrait of bladder cancer: correlation between genomic
aberrations and the DNA damage response. Nature, August 2012.
[70] P V Schoenlein, D W Shen, J T Barrett, I Pastan, and M M Gottesman. Double
132
minute chromosomes carrying the human multidrug resistance 1 and 2 genes are
generated from the dimerization of submicroscopic circular DNAs in colchicine-
selected KB carcinoma cells. Molecular Biology of the Cell, 3(5):507–520, May
1992.
[71] Heidi Schwarzenbach, Dave S B Hoon, and Klaus Pantel. Cell-free nucleic acids as
biomarkers in cancer patients. Nature, 11(6):426–437, May 2011.
[72] M Serrano, G J Hannon, and D Beach. A new regulatory motif in cell-cycle control
causing specific inhibition of cyclin D/CDK4. Nature, 366(6456):704–707, Decem-
ber 1993.
[73] S P Shah. Computational methods for identification of recurrent copy number alter-
ation patterns by array CGH. Cytogenet Genome Res, 123(1-4):343–351, January
2008.
[74] Jonathan S T Sham, Terence C-M Tang, Yan Fang, Li Sun, Lun-Xiu Qin, Qiu-
Liang Wu, Dan Xie, and Xin-Yuan Guan. Recurrent chromosome alterations in
primary ovarian carcinoma in Chinese women. Cancer Genet Cytogenet, 133(1):39–
44, February 2002.
[75] S T Sherry, M H Ward, M Kholodov, J Baker, L Phan, E M Smigielski, and
K Sirotkin. dbSNP: the NCBI database of genetic variation. Nucleic Acids Res,
29(1):308–311, January 2001.
[76] Johan Skog, Tom Wurdinger, Sjoerd van Rijn, Dimphna H Meijer, Laura Gainche,
133
William T Curry, Bob S Carter, Anna M Krichevsky, and Xandra O Breake-
field. Glioblastoma microvesicles transport RNA and proteins that promote tumour
growth and provide diagnostic biomarkers. Nature, 10(12):1470–1476, November
2008.
[77] D J Slamon, B Leyland-Jones, S Shak, H Fuchs, V Paton, A Bajamonde, T Fleming,
W Eiermann, J Wolter, M Pegram, J Baselga, and L Norton. Use of chemother-
apy plus a monoclonal antibody against HER2 for metastatic breast cancer that
overexpresses HER2. N Engl J Med, 344(11):783–792, March 2001.
[78] M L Slovak, J P Ho, M J Pettenati, A Khan, D Douer, S Lal, and S T Traweek.
Localization of amplified MYC gene sequences to double minute chromosomes in
acute myelogenous leukemia. Genes Chromosomes Cancer, 9(1):62–67, January
1994.
[79] T F Smith and M S Waterman. Identification of common molecular subsequences.
Journal of Molecular Biology, 147(1):195–197, March 1981.
[80] T Sørlie, C M Perou, R Tibshirani, T Aas, S Geisler, H Johnsen, T Hastie, M B
Eisen, M van de Rijn, S S Jeffrey, T Thorsen, H Quist, J C Matese, P O Brown,
D Botstein, P Eystein Lønning, and A L Børresen-Dale. Gene expression patterns
of breast carcinomas distinguish tumor subclasses with clinical implications. Proc
Natl Acad Sci USA, 98(19):10869–10874, September 2001.
[81] Therese Sorlie, Robert Tibshirani, Joel Parker, Trevor Hastie, J S Marron, An-
134
drew Nobel, Shibing Deng, Hilde Johnsen, Robert Pesich, Stephanie Geisler, Janos
Demeter, Charles M Perou, Per E Lønning, Patrick O Brown, Anne-Lise Børresen-
Dale, and David Botstein. Repeated observation of breast tumor subtypes in in-
dependent gene expression data sets. Proc Natl Acad Sci USA, 100(14):8418–8423,
July 2003.
[82] Philip J Stephens, Chris D Greenman, Beiyuan Fu, Fengtang Yang, Graham R
Bignell, Laura J Mudie, Erin D Pleasance, King Wai Lau, David Beare, Lucy A
Stebbings, Stuart McLaren, Meng-Lay Lin, David J McBride, Ignacio Varela, Ser-
ena Nik-Zainal, Catherine Leroy, Mingming Jia, Andrew Menzies, Adam P Butler,
Jon W Teague, Michael A Quail, John Burton, Harold Swerdlow, Nigel P Carter,
Laura A Morsberger, Christine Iacobuzio-Donahue, George A Follows, Anthony R
Green, Adrienne M Flanagan, Michael R Stratton, P Andrew Futreal, and Peter J
Campbell. Massive Genomic Rearrangement Acquired in a Single Catastrophic
Event during Cancer Development. Cell, 144(1):27–40, January 2011.
[83] Clelia Tiziana Storlazzi, Thoas Fioretos, Cecilia Surace, Angelo Lonoce, Angela
Mastrorilli, Bodil Strombeck, Pietro D’Addabbo, Francesco Iacovelli, Crescenzio
Minervini, Anna Aventin, Nicole Dastugue, Christa Fonatsch, Anne Hagemeijer,
Martine Jotterand, Dominique Muhlematter, Marina Lafage-Pochitaloff, Florence
Nguyen-Khac, Claudia Schoch, Marilyn L Slovak, Arabella Smith, Francesc Sole,
Nadine Van Roy, Bertil Johansson, and Mariano Rocchi. MYC-containing double
minutes in hematologic malignancies: evidence in favor of the episome model and
135
exclusion of MYC as the target gene. Hum Mol Genet, 15(6):933–942, March 2006.
[84] Clelia Tiziana Storlazzi, Angelo Lonoce, Maria C Guastadisegni, Domenico Trom-
betta, Pietro D’Addabbo, Giulia Daniele, Alberto L’Abbate, Gemma Macchia,
Cecilia Surace, Klaas Kok, Reinhard Ullmann, Stefania Purgato, Orazio Palumbo,
Massimo Carella, Peter F Ambros, and Mariano Rocchi. Gene amplification as dou-
ble minutes or homogeneously staining regions in solid tumors: origin and structure.
Genome Research, 20(9):1198–1206, September 2010.
[85] Aravind Subramanian, Pablo Tamayo, Vamsi K Mootha, Sayan Mukherjee, Ben-
jamin L Ebert, Michael A Gillette, Amanda Paulovich, Scott L Pomeroy, Todd R
Golub, Eric S Lander, and Jill P Mesirov. Gene set enrichment analysis: a
knowledge-based approach for interpreting genome-wide expression profiles. Proc
Natl Acad Sci USA, 102(43):15545–15550, October 2005.
[86] Kenji Takahashi, Takashi Kohno, Rieko Ajima, Hiroki Sasaki, John D Minna,
Toshiyoshi Fujiwara, Noriaki Tanaka, and Jun Yokota. Homozygous deletion and
reduced expression of the DOCK8 gene in human lung cancer. International Jour-
nal of Oncology, 28(2):321.
[87] Scott A Tomlins, Daniel R Rhodes, Sven Perner, Saravana M Dhanasekaran, Rohit
Mehra, Xiao-Wei Sun, Sooryanarayana Varambally, Xuhong Cao, Joelle Tchinda,
Rainer Kuefer, Charles Lee, James E Montie, Rajal B Shah, Kenneth J Pienta,
Mark A Rubin, and Arul M Chinnaiyan. Recurrent fusion of TMPRSS2 and ETS
136
transcription factor genes in prostate cancer. Science, 310(5748):644–648, October
2005.
[88] Kristan E van der Vos, Leonora Balaj, Johan Skog, and Xandra O Breakefield.
Brain tumor microvesicles: insights into intercellular communication in the nervous
system. Cellular and molecular neurobiology, 31(6):949–959, August 2011.
[89] Marileila Varella-Garcia. Stratification of non-small cell lung cancer patients for
therapy with epidermal growth factor receptor inhibitors: the EGFR fluorescence
in situ hybridization assay. Diagnostic Pathology, 1(1):19, 2006.
[90] Imre Vastrik, Peter D’Eustachio, Esther Schmidt, Geeta Joshi-Tope, Gopal
Gopinath, David Croft, Bernard de Bono, Marc Gillespie, Bijay Jassal, Suzanna
Lewis, Lisa Matthews, Guanming Wu, Ewan Birney, and Lincoln Stein. Reactome:
a knowledge base of biologic pathways and processes. Genome Biol, 8(3):R39,
January 2007.
[91] P Vigneri and J Y Wang. Induction of apoptosis in chronic myelogenous leukemia
cells through nuclear entrapment of BCR-ABL tyrosine kinase. Nat Med, 7(2):228–
234, February 2001.
[92] Matthew J Walter, Dong Shen, Li Ding, Jin Shao, Daniel C Koboldt, Ken Chen,
David E Larson, Michael D McLellan, David Dooling, Rachel Abbott, Robert Ful-
ton, Vincent Magrini, Heather Schmidt, Joelle Kalicki-Veizer, Michelle O’Laughlin,
Xian Fan, Marcus Grillot, Sarah Witowski, Sharon Heath, John L Frater, William
137
Eades, Michael Tomasson, Peter Westervelt, John F DiPersio, Daniel C Link,
Elaine R Mardis, Timothy J Ley, Richard K Wilson, and Timothy A Graubert.
Clonal Architecture of Secondary Acute Myeloid Leukemia. N Engl J Med,
366(12):1090–1098, March 2012.
[93] D Wang, X Tian, and Y Jiang. NDRG1/Cap43 overexpression in tumor tissues
and serum from lung cancer patients. Journal of Cancer Research and Clinical
Oncology, June 2012.
[94] J L Wittliff. Steroid-hormone receptors in breast cancer. Cancer, 53(3 Suppl):630–
643, February 1984.
[95] Seiji Yano, Yasuhiko Nishioka, Hisatsugu Goto, and Saburo Sone. Molecular mech-
anisms of angiogenesis in non-small cell lung cancer, and therapeutics targeting
related molecules. Cancer Sci, 94(6):479–485, June 2003.
[96] Lars Zender, Wen Xue, Johannes Zuber, Camile P Semighini, Alexander Kras-
nitz, Beicong Ma, Peggy Zender, Stefan Kubicka, John M Luk, Peter Schirmacher,
W Richard McCombie, Michael Wigler, James Hicks, Gregory J Hannon, Scott
Powers, and Scott W Lowe. An oncogenomics-based in vivo RNAi screen identifies
tumor suppressors in liver cancer. Cell, 135(5):852–864, November 2008.
[97] Zhujiang Zhao, Qingxiang Wu, Jian Cheng, Xuemei Qiu, Jianqiong Zhang, and
Hong Fan. Depletion of DNMT3A Suppressed Cell Proliferation and Restored
138
PTEN in Hepatocellular Carcinoma Cell. Journal of Biomedicine and Biotechnol-
ogy, 2010, May 2010.
[98] Jingchun Zhu, J Zachary Sanborn, Stephen Benz, Christopher Szeto, Fan Hsu,
Robert M Kuhn, Donna Karolchik, John Archie, Marc E Lenburg, Laura J Es-
serman, W James Kent, David Haussler, and Ting Wang. The UCSC Cancer
Genomics Browser. Nat Methods, 6(4):239–240, April 2009.
139